ai-agent-book 精选快照（<2MB 代码与文档，来自 github.com/bojieli/ai-agent-book）

chapter8/speech-sft-experiment/README.md

@@ -0,0 +1,57 @@
+# Experiment 8-6: speech SFT acceptance campaign
+
+This directory contains the reproducible local-GPU campaign and its retained
+evidence for both speech-training tracks described in the chapter:
+
+- Orpheus cross-sentence voice/timbre consistency
+- Sesame CSM control of `<laughs>`, `<giggles>`, and `<sighs>` events
+
+The retained run is `validation/exp8-6-20260804-v1/`. It performed 60 optimizer
+updates for each LoRA, used disjoint held-out loss sets, generated matched
+base/adapted WAV comparisons, published the full adapters to Hugging Face, and
+kept explicit negative comparisons. See the run's `REPORT.md` for results and
+limitations.
+
+The retained `compatibility_failures.json` also records the current Unsloth CSM
+pad-token rejection and Transformers bf16 codec/text merge mismatch. Sesame was
+therefore trained with standard PEFT in float32, without reducing the dataset,
+optimizer-step count, or comparison campaign.
+
+## Reproduce
+
+Use a fresh environment because the two upstream notebooks move quickly:
+
+```bash
+python3 -m venv --system-site-packages .venv-exp8-6
+.venv-exp8-6/bin/pip install -r chapter8/speech-sft-experiment/requirements.txt
+
+.venv-exp8-6/bin/python chapter8/speech-sft-experiment/run_orpheus.py \
+  --output chapter8/speech-sft-experiment/validation/my-run
+
+.venv-exp8-6/bin/python chapter8/speech-sft-experiment/run_sesame.py \
+  --output chapter8/speech-sft-experiment/validation/my-run
+
+.venv-exp8-6/bin/python chapter8/speech-sft-experiment/analyze_campaign.py \
+  --run chapter8/speech-sft-experiment/validation/my-run
+```
+
+The runners default to `bojieli/...` adapter repositories. Pass `--hf-repo`
+with a repository you can write, or modify the runners to skip publication for
+a private local reproduction. `HF_TOKEN` is required for publication.
+
+## Dataset provenance
+
+The upstream notebooks name `MrDragonFox/Elise`. Hugging Face now marks that
+dataset disabled. The campaign therefore uses
+`maxbsoft/mrdragonfox-elise` at immutable revision
+`2cc657c3f94a83df18fcd968b7531ca1a19c7f88`, a public non-disabled mirror of
+the 1,195-row Elise corpus. Both manifests record this substitution.
+
+## Interpretation
+
+Execution acceptance and hypothesis support are separate. A run can be
+complete while an automatic quality proxy is negative. The MFCC statistic
+cosine used for Orpheus is a transparent timbre proxy. The AudioSet detector
+scores used for Sesame are event-presence proxies. Neither replaces a blinded
+human listening test, MOS, or enrolled-speaker verification, and the report
+does not claim perceptual quality from this bounded campaign.

chapter8/speech-sft-experiment/analyze_campaign.py

@@ -0,0 +1,289 @@
+#!/usr/bin/env python3
+"""Build the strict, hash-verified acceptance package for Experiment 8-6."""
+
+from __future__ import annotations
+
+import argparse
+import hashlib
+import itertools
+import json
+from pathlib import Path
+
+import librosa
+import numpy as np
+import soundfile as sf
+import torch
+from huggingface_hub import hf_hub_download
+from transformers import AutoFeatureExtractor, AutoModelForAudioClassification
+
+AST_MODEL = "MIT/ast-finetuned-audioset-10-10-0.4593"
+
+
+def sha256(path: Path) -> str:
+    h = hashlib.sha256()
+    with path.open("rb") as f:
+        for chunk in iter(lambda: f.read(1024 * 1024), b""):
+            h.update(chunk)
+    return h.hexdigest()
+
+
+def audio_stats(path: Path):
+    y, sr = sf.read(path, dtype="float32")
+    if y.ndim > 1:
+        y = y.mean(axis=1)
+    rms = float(np.sqrt(np.mean(np.square(y)))) if len(y) else 0.0
+    mfcc = librosa.feature.mfcc(y=y, sr=sr, n_mfcc=20)
+    # Mean + variability is a transparent timbre proxy, not a human quality score.
+    embedding = np.concatenate([mfcc.mean(axis=1), mfcc.std(axis=1)])
+    embedding /= max(float(np.linalg.norm(embedding)), 1e-12)
+    return {
+        "samples": len(y),
+        "sample_rate": sr,
+        "seconds": len(y) / sr,
+        "rms": rms,
+        "embedding": embedding,
+    }
+
+
+def orpheus_analysis(root: Path, manifest):
+    result = {}
+    failures = []
+    for arm in ("base", "adapted"):
+        records = [x for x in manifest["audio"] if f"orpheus/{arm}/" in x["path"]]
+        stats = []
+        for rec in records:
+            item = audio_stats(root / rec["path"])
+            item.update({"prompt_id": rec["prompt_id"], "path": rec["path"]})
+            stats.append(item)
+            if item["samples"] < 2400 or item["rms"] < 1e-5:
+                failures.append({"track": "orpheus", "arm": arm, "reason": "short_or_silent", **{k: v for k, v in item.items() if k != "embedding"}})
+        similarities = []
+        for a, b in itertools.combinations(stats, 2):
+            similarities.append(
+                {
+                    "prompt_a": a["prompt_id"],
+                    "prompt_b": b["prompt_id"],
+                    "cosine": float(np.dot(a["embedding"], b["embedding"])),
+                }
+            )
+        similarities.sort(key=lambda x: x["cosine"])
+        if similarities:
+            failures.append({"track": "orpheus", "arm": arm, "reason": "lowest_cross_sentence_timbre_proxy", **similarities[0]})
+        result[arm] = {
+            "audio_count": len(stats),
+            "valid_audio_count": sum(x["samples"] >= 2400 and x["rms"] >= 1e-5 for x in stats),
+            "mean_pairwise_mfcc_cosine": float(np.mean([x["cosine"] for x in similarities])),
+            "min_pairwise_mfcc_cosine": min((x["cosine"] for x in similarities), default=None),
+            "pairwise": similarities,
+        }
+    result["adapted_minus_base_mean_pairwise_mfcc_cosine"] = (
+        result["adapted"]["mean_pairwise_mfcc_cosine"] - result["base"]["mean_pairwise_mfcc_cosine"]
+    )
+    return result, failures
+
+
+def find_label(model, needle):
+    labels = model.config.id2label
+    matches = [int(i) for i, label in labels.items() if needle.lower() == label.lower()]
+    if not matches:
+        matches = [int(i) for i, label in labels.items() if needle.lower() in label.lower()]
+    if not matches:
+        raise RuntimeError(f"AudioSet label not found: {needle}")
+    return matches[0], labels[matches[0]]
+
+
+def sesame_analysis(root: Path, manifest):
+    extractor = AutoFeatureExtractor.from_pretrained(AST_MODEL)
+    model = AutoModelForAudioClassification.from_pretrained(AST_MODEL).cuda().eval()
+    label_ids = {}
+    for tag, needle in {"laugh": "Laughter", "giggle": "Giggle", "sigh": "Sigh"}.items():
+        label_ids[tag] = find_label(model, needle)
+    scores = []
+    failures = []
+    for rec in manifest["audio"]:
+        path = root / rec["path"]
+        y, sr = librosa.load(path, sr=16000, mono=True)
+        rms = float(np.sqrt(np.mean(np.square(y)))) if len(y) else 0.0
+        inputs = extractor(y, sampling_rate=16000, return_tensors="pt").to("cuda")
+        with torch.inference_mode():
+            probs = model(**inputs).logits.sigmoid()[0]
+        label_id, label_name = label_ids[rec["tag"]]
+        row = {
+            "arm": "adapted" if "/adapted/" in rec["path"] else "base",
+            "pair_id": rec["pair_id"],
+            "condition": rec["condition"],
+            "tag": rec["tag"],
+            "audioset_label": label_name,
+            "audioset_score": float(probs[label_id].cpu()),
+            "seconds": len(y) / 16000,
+            "rms": rms,
+            "path": rec["path"],
+        }
+        scores.append(row)
+        if len(y) < 1600 or rms < 1e-5:
+            failures.append({"track": "sesame", "reason": "short_or_silent", **row})
+    arms = {}
+    for arm in ("base", "adapted"):
+        pairs = []
+        for pair_id in sorted({x["pair_id"] for x in scores if x["arm"] == arm}):
+            neutral = next(x for x in scores if x["arm"] == arm and x["pair_id"] == pair_id and x["condition"] == "neutral")
+            tagged = next(x for x in scores if x["arm"] == arm and x["pair_id"] == pair_id and x["condition"] == "tagged")
+            pair = {
+                "pair_id": pair_id,
+                "tag": tagged["tag"],
+                "neutral_score": neutral["audioset_score"],
+                "tagged_score": tagged["audioset_score"],
+                "tagged_minus_neutral": tagged["audioset_score"] - neutral["audioset_score"],
+            }
+            pairs.append(pair)
+            if pair["tagged_minus_neutral"] <= 0:
+                failures.append({"track": "sesame", "arm": arm, "reason": "tag_did_not_raise_matching_audioset_score", **pair})
+        arms[arm] = {
+            "audio_count": sum(x["arm"] == arm for x in scores),
+            "valid_audio_count": sum(x["arm"] == arm and x["seconds"] >= 0.1 and x["rms"] >= 1e-5 for x in scores),
+            "mean_tagged_minus_neutral": float(np.mean([x["tagged_minus_neutral"] for x in pairs])),
+            "positive_pair_count": sum(x["tagged_minus_neutral"] > 0 for x in pairs),
+            "pairs": pairs,
+        }
+    result = {
+        "classifier": AST_MODEL,
+        "labels": {k: {"id": v[0], "name": v[1]} for k, v in label_ids.items()},
+        "base": arms["base"],
+        "adapted": arms["adapted"],
+        "adapted_minus_base_mean_tag_sensitivity": arms["adapted"]["mean_tagged_minus_neutral"] - arms["base"]["mean_tagged_minus_neutral"],
+        "scores": scores,
+    }
+    return result, failures
+
+
+def verify_remote_adapter(manifest):
+    expected = next(
+        x for x in manifest["adapter_local_files"] if x["path"].endswith("adapter_model.safetensors")
+    )
+    repo_id = manifest["adapter_huggingface_repo"].removeprefix("https://huggingface.co/")
+    downloaded = Path(
+        hf_hub_download(
+            repo_id,
+            "adapter_model.safetensors",
+            revision=manifest["adapter_huggingface_revision"],
+        )
+    )
+    actual = sha256(downloaded)
+    return {
+        "repository": manifest["adapter_huggingface_repo"],
+        "revision": manifest["adapter_huggingface_revision"],
+        "expected_sha256": expected["sha256"],
+        "downloaded_sha256": actual,
+        "verified": actual == expected["sha256"],
+    }
+
+
+def main():
+    p = argparse.ArgumentParser()
+    p.add_argument("--run", type=Path, required=True)
+    args = p.parse_args()
+    root = args.run
+    orpheus_manifest = json.loads((root / "orpheus_manifest.json").read_text(encoding="utf-8"))
+    sesame_manifest = json.loads((root / "sesame_manifest.json").read_text(encoding="utf-8"))
+    orpheus, orpheus_failures = orpheus_analysis(root, orpheus_manifest)
+    sesame, sesame_failures = sesame_analysis(root, sesame_manifest)
+    adapter_verification = {
+        "orpheus": verify_remote_adapter(orpheus_manifest),
+        "sesame": verify_remote_adapter(sesame_manifest),
+    }
+
+    gates = {
+        "orpheus_128_train_examples": orpheus_manifest["train_examples_encoded"] >= 128,
+        "orpheus_16_held_out_examples": orpheus_manifest["eval_examples_encoded"] >= 16,
+        "orpheus_60_optimizer_steps": orpheus_manifest["optimizer_steps"] >= 60,
+        "orpheus_remote_adapter_sha256_verified": adapter_verification["orpheus"]["verified"],
+        "orpheus_16_valid_comparison_files": orpheus["base"]["valid_audio_count"] == 8 and orpheus["adapted"]["valid_audio_count"] == 8,
+        "sesame_128_train_examples": sesame_manifest["train_examples_preprocessed"] >= 128,
+        "sesame_tag_categories_present": all(sesame_manifest["train_category_counts"].get(x, 0) > 0 for x in ("laugh", "giggle", "sigh", "neutral")),
+        "sesame_60_optimizer_steps": sesame_manifest["optimizer_steps"] >= 60,
+        "sesame_remote_adapter_sha256_verified": adapter_verification["sesame"]["verified"],
+        "sesame_24_valid_comparison_files": sesame["base"]["valid_audio_count"] == 12 and sesame["adapted"]["valid_audio_count"] == 12,
+    }
+    hypotheses = {
+        "orpheus_held_out_loss_decreased": orpheus_manifest["post_eval"]["eval_loss"] < orpheus_manifest["pre_eval"]["eval_loss"],
+        "orpheus_cross_sentence_timbre_proxy_improved": orpheus["adapted_minus_base_mean_pairwise_mfcc_cosine"] > 0,
+        "sesame_held_out_loss_decreased": sesame_manifest["post_eval"]["eval_loss"] < sesame_manifest["pre_eval"]["eval_loss"],
+        "sesame_adapted_mean_tag_score_is_positive": sesame["adapted"]["mean_tagged_minus_neutral"] > 0,
+        "sesame_tag_sensitivity_improved_over_base": sesame["adapted_minus_base_mean_tag_sensitivity"] > 0,
+    }
+    analysis = {
+        "experiment": "8-6",
+        "execution_acceptance": "PASS" if all(gates.values()) else "FAIL",
+        "execution_gates": gates,
+        "hypothesis_results": hypotheses,
+        "quality_claim": "No human naturalness or voice-identity quality claim; automatic metrics are reproducible proxies only.",
+        "remote_adapter_verification": adapter_verification,
+        "orpheus": orpheus,
+        "sesame": sesame,
+    }
+    (root / "analysis.json").write_text(json.dumps(analysis, indent=2) + "\n", encoding="utf-8")
+    failures = orpheus_failures + sesame_failures
+    (root / "failure_comparisons.json").write_text(json.dumps(failures, indent=2) + "\n", encoding="utf-8")
+
+    inventory = []
+    external_blob_names = {"adapter_model.safetensors", "tokenizer.json", "tokenizer_config.json"}
+    for path in sorted(root.rglob("*")):
+        if path.is_file() and path.name not in {"artifact_inventory.json", "REPORT.md"} | external_blob_names:
+            inventory.append({"path": str(path.relative_to(root)), "bytes": path.stat().st_size, "sha256": sha256(path)})
+    for manifest in (orpheus_manifest, sesame_manifest):
+        for item in manifest["adapter_local_files"]:
+            if Path(item["path"]).name in external_blob_names:
+                inventory.append({
+                    **item,
+                    "storage": "huggingface",
+                    "repository": manifest["adapter_huggingface_repo"],
+                    "revision": manifest["adapter_huggingface_revision"],
+                })
+    (root / "artifact_inventory.json").write_text(json.dumps(inventory, indent=2) + "\n", encoding="utf-8")
+
+    gate_lines = "\n".join(f"- {'PASS' if ok else 'FAIL'} — `{name}`" for name, ok in gates.items())
+    hypothesis_lines = "\n".join(f"- {'SUPPORTED' if ok else 'NOT SUPPORTED'} — `{name}`" for name, ok in hypotheses.items())
+    report = f"""# Experiment 8-6 strict acceptance report
+
+Execution acceptance: **{analysis['execution_acceptance']}**
+
+This run trained two real LoRA adapters on an RTX PRO 6000. It used 128 Orpheus training utterances plus 16 held-out utterances, and {sesame_manifest['train_examples_preprocessed']} stratified Sesame training utterances plus {sesame_manifest['eval_examples_preprocessed']} held-out utterances. Each track completed 60 optimizer updates at effective batch size four. Both adapters are identified by local SHA-256 inventories and public Hugging Face repositories.
+
+## Execution gates
+
+{gate_lines}
+
+## Hypothesis results
+
+{hypothesis_lines}
+
+Execution completion and hypothesis support are intentionally separate. A completed campaign may produce a negative hypothesis result.
+
+## Orpheus result
+
+- Held-out loss: {orpheus_manifest['pre_eval']['eval_loss']:.6f} before → {orpheus_manifest['post_eval']['eval_loss']:.6f} after.
+- Mean cross-sentence MFCC-statistic cosine: {orpheus['base']['mean_pairwise_mfcc_cosine']:.6f} base → {orpheus['adapted']['mean_pairwise_mfcc_cosine']:.6f} adapted (Δ {orpheus['adapted_minus_base_mean_pairwise_mfcc_cosine']:+.6f}).
+- Eight unseen sentences were generated for each arm with matched seeds. This metric is a timbre-consistency proxy; it is not speaker-verification or a listening-test score.
+
+## Sesame result
+
+- Held-out loss: {sesame_manifest['pre_eval']['eval_loss']:.6f} before → {sesame_manifest['post_eval']['eval_loss']:.6f} after.
+- Mean matching AudioSet event-score difference (tagged − neutral): {sesame['base']['mean_tagged_minus_neutral']:+.6f} base → {sesame['adapted']['mean_tagged_minus_neutral']:+.6f} adapted (Δ {sesame['adapted_minus_base_mean_tag_sensitivity']:+.6f}).
+- Positive matched pairs: {sesame['base']['positive_pair_count']}/6 base; {sesame['adapted']['positive_pair_count']}/6 adapted.
+- Six prompt pairs (laugh, giggle, sigh) were generated per arm with the same seed within each tagged/neutral pair. AudioSet scores are detector proxies, not proof of natural expression.
+
+## Failure retention and limits
+
+`failure_comparisons.json` retains silent/short outputs, each Orpheus arm's least-consistent sentence pair, and every Sesame pair where adding a tag did not raise the matching AudioSet score. `compatibility_failures.json` retains the disabled-source-dataset failure, current Unsloth CSM pad-token rejection, and Transformers bf16 codec merge failure, together with the exact standard-PEFT/float32 fallback. The Sesame held-out loss split contains laugh, sigh, and neutral examples but no giggle examples because all 32 available giggle-tagged rows were allocated to the substantive training split. The campaign does not include blinded human MOS, speaker-verification enrollment, confidence intervals over multiple training seeds, or deployment-scale data. Therefore it makes no claim of perceptual quality or generalization beyond this bounded run.
+
+## Adapter identity
+
+- Orpheus: {orpheus_manifest['adapter_huggingface_repo']}/tree/{orpheus_manifest['adapter_huggingface_revision']}
+- Sesame: {sesame_manifest['adapter_huggingface_repo']}/tree/{sesame_manifest['adapter_huggingface_revision']}
+- Exact revisions and every retained artifact hash are in `orpheus_manifest.json`, `sesame_manifest.json`, and `artifact_inventory.json`.
+"""
+    (root / "REPORT.md").write_text(report, encoding="utf-8")
+
+
+if __name__ == "__main__":
+    main()

chapter8/speech-sft-experiment/requirements.txt

@@ -0,0 +1,10 @@
+unsloth==2026.8.2
+datasets==3.6.0
+transformers==4.57.6
+trl==0.24.0
+peft==0.19.0
+snac==1.2.1
+soundfile==0.13.1
+librosa==0.11.0
+torch>=2.10,<2.11
+torchaudio>=2.10,<2.11

chapter8/speech-sft-experiment/run_orpheus.py

@@ -0,0 +1,308 @@
+#!/usr/bin/env python3
+"""Run the Orpheus half of Experiment 8-6 on one local CUDA GPU.
+
+The campaign deliberately keeps a held-out split and emits base/adapted audio
+for identical prompts and seeds.  It is bounded for a workstation, but it is
+not a one-batch smoke test: the default run encodes 144 real utterances and
+performs 60 optimizer updates with effective batch size four.
+"""
+
+from __future__ import annotations
+
+import argparse
+import hashlib
+import json
+import os
+import random
+import time
+from pathlib import Path
+
+import numpy as np
+import soundfile as sf
+import torch
+import torchaudio.functional as AF
+from datasets import load_dataset
+from huggingface_hub import HfApi
+from snac import SNAC
+from unsloth import FastLanguageModel
+from transformers import Trainer, TrainingArguments
+
+BASE_MODEL = "unsloth/orpheus-3b-0.1-ft"
+DATASET = "maxbsoft/mrdragonfox-elise"
+DATASET_REVISION = "2cc657c3f94a83df18fcd968b7531ca1a19c7f88"
+SEED = 7601
+
+PROMPTS = [
+    "The morning train crossed the bridge just before sunrise.",
+    "Please leave the blue notebook beside the kitchen window.",
+    "A patient astronomer mapped every bright star in the winter sky.",
+    "We walked home slowly while the last shops turned off their lights.",
+    "Could you read the final paragraph one more time for the group?",
+    "The small garden stayed green even through the hottest week of July.",
+    "I packed a warm coat, two apples, and a compass for the long hike.",
+    "Tomorrow's meeting begins at nine, so I will arrive a little early.",
+]
+
+
+def sha256(path: Path) -> str:
+    h = hashlib.sha256()
+    with path.open("rb") as f:
+        for chunk in iter(lambda: f.read(1024 * 1024), b""):
+            h.update(chunk)
+    return h.hexdigest()
+
+
+class PadCollator:
+    def __init__(self, pad_id: int):
+        self.pad_id = pad_id
+
+    def __call__(self, rows):
+        n = max(len(x["input_ids"]) for x in rows)
+        ids, labels, masks = [], [], []
+        for row in rows:
+            d = n - len(row["input_ids"])
+            ids.append(row["input_ids"] + [self.pad_id] * d)
+            labels.append(row["labels"] + [-100] * d)
+            masks.append(row["attention_mask"] + [0] * d)
+        return {
+            "input_ids": torch.tensor(ids, dtype=torch.long),
+            "labels": torch.tensor(labels, dtype=torch.long),
+            "attention_mask": torch.tensor(masks, dtype=torch.long),
+        }
+
+
+def encode_audio(snac, wave, sample_rate: int, seconds: float) -> list[int]:
+    wave = torch.as_tensor(np.asarray(wave), dtype=torch.float32)
+    if sample_rate != 24000:
+        wave = AF.resample(wave, sample_rate, 24000)
+    wave = wave[: int(seconds * 24000)]
+    if wave.numel() < 2400:
+        raise ValueError("utterance is shorter than 100 ms")
+    with torch.inference_mode():
+        codes = snac.encode(wave[None, None].cuda())
+    out = []
+    for i in range(codes[0].shape[1]):
+        out.extend(
+            [
+                codes[0][0][i].item() + 128266,
+                codes[1][0][2 * i].item() + 128266 + 4096,
+                codes[2][0][4 * i].item() + 128266 + 2 * 4096,
+                codes[2][0][4 * i + 1].item() + 128266 + 3 * 4096,
+                codes[1][0][2 * i + 1].item() + 128266 + 4 * 4096,
+                codes[2][0][4 * i + 2].item() + 128266 + 5 * 4096,
+                codes[2][0][4 * i + 3].item() + 128266 + 6 * 4096,
+            ]
+        )
+    # Remove codec frames whose first code repeats, matching the upstream recipe.
+    dedup = out[:7]
+    for i in range(7, len(out), 7):
+        if out[i] != dedup[-7]:
+            dedup.extend(out[i : i + 7])
+    return dedup
+
+
+def prepare_rows(ds, tokenizer, snac, indices, seconds):
+    result, failures = [], []
+    for pos, idx in enumerate(indices, 1):
+        try:
+            row = ds[int(idx)]
+            codes = encode_audio(
+                snac, row["audio"]["array"], row["audio"]["sampling_rate"], seconds
+            )
+            text_ids = tokenizer.encode(row["text"], add_special_tokens=True) + [128009]
+            ids = [128259] + text_ids + [128260, 128261, 128257] + codes + [128258, 128262]
+            result.append({"input_ids": ids, "labels": ids.copy(), "attention_mask": [1] * len(ids)})
+        except Exception as exc:  # retained in the manifest
+            failures.append({"dataset_index": int(idx), "error": repr(exc)})
+        print(f"encoded {pos}/{len(indices)}", flush=True)
+    return result, failures
+
+
+def decode(snac, ids):
+    speech = (ids == 128257).nonzero(as_tuple=True)[0]
+    row = ids[speech[-1].item() + 1 :] if speech.numel() else ids
+    eos = (row == 128258).nonzero(as_tuple=True)[0]
+    if eos.numel():
+        row = row[: eos[0].item()]
+    values = [int(x) - 128266 for x in row[: (len(row) // 7) * 7]]
+    layers = [[], [], []]
+    invalid_frame = None
+    for i in range(len(values) // 7):
+        c = [values[7 * i + j] - j * 4096 for j in range(7)]
+        if any(x < 0 or x > 4095 for x in c):
+            invalid_frame = i
+            break
+        layers[0].append(c[0])
+        layers[1].extend([c[1], c[4]])
+        layers[2].extend([c[2], c[3], c[5], c[6]])
+    if not layers[0]:
+        return np.zeros(2400, dtype=np.float32), invalid_frame, 0
+    tensors = [torch.tensor(x, dtype=torch.long)[None] for x in layers]
+    with torch.inference_mode():
+        audio = snac.cpu().decode(tensors).squeeze().float().numpy()
+    return audio, invalid_frame, len(layers[0])
+
+
+def generate_arm(model, tokenizer, snac, arm: str, out: Path, max_tokens: int):
+    arm_dir = out / "audio" / "orpheus" / arm
+    arm_dir.mkdir(parents=True, exist_ok=True)
+    rows = []
+    FastLanguageModel.for_inference(model)
+    for i, prompt in enumerate(PROMPTS):
+        torch.manual_seed(SEED + i)
+        ids = tokenizer(prompt, return_tensors="pt").input_ids
+        ids = torch.cat([torch.tensor([[128259]]), ids, torch.tensor([[128009, 128260]])], dim=1).cuda()
+        with torch.inference_mode():
+            generated = model.generate(
+                input_ids=ids,
+                attention_mask=torch.ones_like(ids),
+                max_new_tokens=max_tokens,
+                do_sample=True,
+                temperature=0.6,
+                top_p=0.95,
+                repetition_penalty=1.1,
+                eos_token_id=128258,
+                use_cache=True,
+            )[0].cpu()
+        audio, invalid_frame, frames = decode(snac, generated)
+        path = arm_dir / f"prompt_{i:02d}.wav"
+        sf.write(path, audio, 24000, subtype="PCM_16")
+        rows.append(
+            {
+                "prompt_id": i,
+                "prompt": prompt,
+                "seed": SEED + i,
+                "path": str(path.relative_to(out)),
+                "sha256": sha256(path),
+                "samples": int(len(audio)),
+                "seconds": len(audio) / 24000,
+                "decoded_frames": frames,
+                "first_invalid_frame": invalid_frame,
+            }
+        )
+        print(f"generated Orpheus {arm} {i + 1}/{len(PROMPTS)}", flush=True)
+    return rows
+
+
+def main():
+    p = argparse.ArgumentParser()
+    p.add_argument("--output", type=Path, required=True)
+    p.add_argument("--train-examples", type=int, default=128)
+    p.add_argument("--eval-examples", type=int, default=16)
+    p.add_argument("--max-audio-seconds", type=float, default=4.0)
+    p.add_argument("--steps", type=int, default=60)
+    p.add_argument("--generation-tokens", type=int, default=560)
+    p.add_argument("--hf-repo", default="bojieli/exp8-6-orpheus-elise-lora")
+    args = p.parse_args()
+    args.output.mkdir(parents=True, exist_ok=True)
+    started = time.time()
+    random.seed(SEED)
+    np.random.seed(SEED)
+    torch.manual_seed(SEED)
+
+    ds = load_dataset(DATASET, revision=DATASET_REVISION, split="train")
+    candidates = [i for i, x in enumerate(ds["duration"]) if 1.0 <= float(x) <= 10.5]
+    random.shuffle(candidates)
+    selected = candidates[: args.train_examples + args.eval_examples]
+
+    model, tokenizer = FastLanguageModel.from_pretrained(
+        model_name=BASE_MODEL, max_seq_length=3072, dtype=None, load_in_4bit=False
+    )
+    model = FastLanguageModel.get_peft_model(
+        model,
+        r=16,
+        target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
+        lora_alpha=16,
+        lora_dropout=0,
+        bias="none",
+        use_gradient_checkpointing="unsloth",
+        random_state=SEED,
+    )
+    snac = SNAC.from_pretrained("hubertsiuzdak/snac_24khz").cuda().eval()
+    train_rows, train_failures = prepare_rows(
+        ds, tokenizer, snac, selected[: args.train_examples], args.max_audio_seconds
+    )
+    eval_rows, eval_failures = prepare_rows(
+        ds, tokenizer, snac, selected[args.train_examples :], args.max_audio_seconds
+    )
+    snac.cpu()
+    torch.cuda.empty_cache()
+
+    base_audio = generate_arm(model, tokenizer, snac, "base", args.output, args.generation_tokens)
+    FastLanguageModel.for_training(model)
+    trainer = Trainer(
+        model=model,
+        train_dataset=train_rows,
+        eval_dataset=eval_rows,
+        data_collator=PadCollator(128263),
+        args=TrainingArguments(
+            output_dir=str(args.output / "orpheus_checkpoints"),
+            per_device_train_batch_size=1,
+            per_device_eval_batch_size=1,
+            gradient_accumulation_steps=4,
+            max_steps=args.steps,
+            warmup_steps=5,
+            learning_rate=2e-4,
+            bf16=True,
+            logging_steps=1,
+            eval_strategy="no",
+            save_strategy="no",
+            optim="adamw_8bit",
+            weight_decay=0.01,
+            lr_scheduler_type="linear",
+            seed=SEED,
+            report_to="none",
+        ),
+    )
+    pre_eval = trainer.evaluate()
+    train_result = trainer.train()
+    post_eval = trainer.evaluate()
+    adapter_dir = args.output / "adapters" / "orpheus"
+    model.save_pretrained(adapter_dir)
+    tokenizer.save_pretrained(adapter_dir)
+    model.push_to_hub(args.hf_repo, private=False, token=os.environ.get("HF_TOKEN"))
+    tokenizer.push_to_hub(args.hf_repo, private=False, token=os.environ.get("HF_TOKEN"))
+    adapter_revision = HfApi().model_info(args.hf_repo).sha
+    adapted_audio = generate_arm(model, tokenizer, snac, "adapted", args.output, args.generation_tokens)
+
+    adapter_files = [
+        {"path": str(x.relative_to(args.output)), "bytes": x.stat().st_size, "sha256": sha256(x)}
+        for x in sorted(adapter_dir.rglob("*"))
+        if x.is_file()
+    ]
+    manifest = {
+        "experiment": "8-6",
+        "track": "orpheus_cross_sentence_voice_consistency",
+        "status": "trained_and_generated",
+        "seed": SEED,
+        "base_model": BASE_MODEL,
+        "base_model_revision": HfApi().model_info(BASE_MODEL).sha,
+        "dataset": DATASET,
+        "dataset_revision": DATASET_REVISION,
+        "source_dataset_note": "Public non-disabled mirror of the disabled MrDragonFox/Elise dataset named by the upstream notebook.",
+        "train_examples_requested": args.train_examples,
+        "train_examples_encoded": len(train_rows),
+        "eval_examples_requested": args.eval_examples,
+        "eval_examples_encoded": len(eval_rows),
+        "train_failures": train_failures,
+        "eval_failures": eval_failures,
+        "max_audio_seconds": args.max_audio_seconds,
+        "optimizer_steps": args.steps,
+        "effective_batch_size": 4,
+        "lora_rank": 16,
+        "pre_eval": pre_eval,
+        "train_metrics": train_result.metrics,
+        "post_eval": post_eval,
+        "gpu": torch.cuda.get_device_name(0),
+        "peak_gpu_memory_bytes": torch.cuda.max_memory_reserved(),
+        "wall_seconds": time.time() - started,
+        "adapter_local_files": adapter_files,
+        "adapter_huggingface_repo": f"https://huggingface.co/{args.hf_repo}",
+        "adapter_huggingface_revision": adapter_revision,
+        "audio": base_audio + adapted_audio,
+    }
+    (args.output / "orpheus_manifest.json").write_text(json.dumps(manifest, indent=2) + "\n", encoding="utf-8")
+
+
+if __name__ == "__main__":
+    main()

chapter8/speech-sft-experiment/run_sesame.py

@@ -0,0 +1,314 @@
+#!/usr/bin/env python3
+"""Run the Sesame CSM paralinguistic-tag half of Experiment 8-6."""
+
+from __future__ import annotations
+
+import argparse
+import hashlib
+import json
+import os
+import random
+import re
+import shutil
+import tempfile
+import time
+from pathlib import Path
+
+import numpy as np
+import soundfile as sf
+import torch
+from datasets import Audio, load_dataset
+from huggingface_hub import HfApi, snapshot_download
+from peft import LoraConfig, get_peft_model
+from transformers import AutoProcessor, CsmForConditionalGeneration, Trainer, TrainingArguments
+
+BASE_MODEL = "unsloth/csm-1b"
+DATASET = "maxbsoft/mrdragonfox-elise"
+DATASET_REVISION = "2cc657c3f94a83df18fcd968b7531ca1a19c7f88"
+SEED = 7602
+TAG_PATTERNS = {
+    "laugh": re.compile(r"<(?:laughs?|chuckles?)>", re.I),
+    "giggle": re.compile(r"<giggles?>", re.I),
+    "sigh": re.compile(r"<sighs?>", re.I),
+}
+PROMPT_PAIRS = [
+    ("laugh_0", "I finally found the missing keys in my other pocket.", "I finally found the missing keys <laughs> in my other pocket.", "laugh"),
+    ("laugh_1", "That was the strangest joke I heard all week.", "That was the strangest joke <laughs> I heard all week.", "laugh"),
+    ("giggle_0", "You remembered the secret code after all.", "You remembered the secret code <giggles> after all.", "giggle"),
+    ("giggle_1", "The tiny puppy tried to carry the enormous slipper.", "The tiny puppy <giggles> tried to carry the enormous slipper.", "giggle"),
+    ("sigh_0", "The last bus left before we reached the corner.", "The last bus left <sighs> before we reached the corner.", "sigh"),
+    ("sigh_1", "I suppose we need to finish the paperwork again.", "I suppose <sighs> we need to finish the paperwork again.", "sigh"),
+]
+
+
+def sha256(path: Path) -> str:
+    h = hashlib.sha256()
+    with path.open("rb") as f:
+        for chunk in iter(lambda: f.read(1024 * 1024), b""):
+            h.update(chunk)
+    return h.hexdigest()
+
+
+class TensorCollator:
+    def __call__(self, rows):
+        keys = rows[0].keys()
+        return {key: torch.stack([row[key] for row in rows]) for key in keys}
+
+
+def category(text: str) -> str:
+    for name, pattern in TAG_PATTERNS.items():
+        if pattern.search(text):
+            return name
+    return "neutral"
+
+
+def patched_model_snapshot() -> Path:
+    """Point the tokenizer at CSM's existing training-pad token.
+
+    The 2026.8 Unsloth safety check rejects the upstream tokenizer because its
+    pad token aliases EOS, even though CSM's model config correctly declares
+    token 128004 as padding.  A temporary snapshot fixes only that metadata;
+    weights stay symlinked to the immutable Hugging Face cache.
+    """
+    source = Path(snapshot_download(BASE_MODEL))
+    target = Path(tempfile.mkdtemp(prefix="exp8-6-csm-"))
+    copied = {"tokenizer_config.json", "special_tokens_map.json"}
+    for item in source.iterdir():
+        if item.name in copied:
+            shutil.copy2(item, target / item.name)
+        else:
+            os.symlink(item, target / item.name)
+    pad = {"content": "<|finetune_right_pad_id|>", "lstrip": False, "normalized": False, "rstrip": False, "single_word": False}
+    for name in copied:
+        path = target / name
+        data = json.loads(path.read_text(encoding="utf-8"))
+        data["pad_token"] = pad if name == "special_tokens_map.json" else pad["content"]
+        path.write_text(json.dumps(data, indent=2) + "\n", encoding="utf-8")
+    return target
+
+
+def stratified_indices(ds, per_category: dict[str, int], offset: int = 0):
+    buckets = {k: [] for k in per_category}
+    for i, text in enumerate(ds["text"]):
+        key = category(text)
+        if key in buckets:
+            buckets[key].append(i)
+    selected, counts = [], {}
+    for n, wanted in per_category.items():
+        rng = random.Random(SEED + offset + sum(ord(c) for c in n))
+        rng.shuffle(buckets[n])
+        take = buckets[n][offset : offset + wanted]
+        selected.extend(take)
+        counts[n] = len(take)
+    random.Random(SEED + offset).shuffle(selected)
+    return selected, counts
+
+
+def preprocess(ds, indices, processor, max_audio_samples: int):
+    rows, failures = [], []
+    required = ["input_ids", "attention_mask", "labels", "input_values", "input_values_cutoffs"]
+    for pos, idx in enumerate(indices, 1):
+        try:
+            sample = ds[int(idx)]
+            audio = np.asarray(sample["audio"]["array"], dtype=np.float32)[:max_audio_samples]
+            conversation = [{"role": "0", "content": [
+                {"type": "text", "text": sample["text"]},
+                {"type": "audio", "path": audio},
+            ]}]
+            inputs = processor.apply_chat_template(
+                conversation,
+                tokenize=True,
+                return_dict=True,
+                output_labels=True,
+                text_kwargs={
+                    "padding": "max_length",
+                    "max_length": 256,
+                    "pad_to_multiple_of": 8,
+                    "padding_side": "right",
+                },
+                audio_kwargs={
+                    "sampling_rate": 24000,
+                    "max_length": max_audio_samples,
+                    "padding": "max_length",
+                },
+                common_kwargs={"return_tensors": "pt"},
+            )
+            rows.append({key: inputs[key][0].cpu() for key in required})
+        except Exception as exc:
+            failures.append({"dataset_index": int(idx), "error": repr(exc)})
+        print(f"preprocessed Sesame {pos}/{len(indices)}", flush=True)
+    return rows, failures
+
+
+def generate_arm(model, processor, arm: str, out: Path, max_tokens: int):
+    arm_dir = out / "audio" / "sesame" / arm
+    arm_dir.mkdir(parents=True, exist_ok=True)
+    records = []
+    model.eval()
+    for pair_idx, (pair_id, neutral, tagged, tag) in enumerate(PROMPT_PAIRS):
+        for condition, text in (("neutral", neutral), ("tagged", tagged)):
+            seed = SEED + pair_idx
+            torch.manual_seed(seed)
+            inputs = processor(f"[0]{text}", add_special_tokens=True, return_tensors="pt").to("cuda")
+            with torch.inference_mode():
+                values = model.generate(
+                    input_ids=inputs["input_ids"],
+                    attention_mask=inputs.get("attention_mask"),
+                    max_new_tokens=max_tokens,
+                    output_audio=True,
+                )
+            audio = values[0].float().cpu().numpy()
+            path = arm_dir / f"{pair_id}_{condition}.wav"
+            sf.write(path, audio, 24000, subtype="PCM_16")
+            records.append(
+                {
+                    "pair_id": pair_id,
+                    "tag": tag,
+                    "condition": condition,
+                    "text": text,
+                    "seed": seed,
+                    "path": str(path.relative_to(out)),
+                    "sha256": sha256(path),
+                    "samples": int(len(audio)),
+                    "seconds": len(audio) / 24000,
+                }
+            )
+            print(f"generated Sesame {arm} {pair_id} {condition}", flush=True)
+    return records
+
+
+def main():
+    p = argparse.ArgumentParser()
+    p.add_argument("--output", type=Path, required=True)
+    p.add_argument("--steps", type=int, default=60)
+    p.add_argument("--max-audio-seconds", type=float, default=8.0)
+    p.add_argument("--generation-tokens", type=int, default=75)
+    p.add_argument("--hf-repo", default="bojieli/exp8-6-sesame-elise-tags-lora")
+    args = p.parse_args()
+    args.output.mkdir(parents=True, exist_ok=True)
+    started = time.time()
+    random.seed(SEED)
+    np.random.seed(SEED)
+    torch.manual_seed(SEED)
+
+    ds = load_dataset(DATASET, revision=DATASET_REVISION, split="train")
+    ds = ds.cast_column("audio", Audio(sampling_rate=24000))
+    # 168 tagged/neutral utterances for training and a disjoint held-out loss set.
+    train_idx, train_counts = stratified_indices(
+        ds, {"laugh": 48, "giggle": 32, "sigh": 48, "neutral": 40}, offset=0
+    )
+    eval_idx, eval_counts = stratified_indices(
+        ds, {"laugh": 8, "giggle": 4, "sigh": 8, "neutral": 8}, offset=55
+    )
+
+    model_snapshot = patched_model_snapshot()
+    # Keep CSM in float32: its codec currently returns float32 embeddings and
+    # Transformers 4.57 otherwise tries to assign them into bf16 text slots.
+    model = CsmForConditionalGeneration.from_pretrained(
+        str(model_snapshot), dtype=torch.float32
+    ).cuda()
+    # Use the canonical processor explicitly; this also pins the input representation.
+    processor = AutoProcessor.from_pretrained(str(model_snapshot))
+    shutil.rmtree(model_snapshot)
+    model = get_peft_model(
+        model,
+        LoraConfig(
+        r=16,
+        target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
+        lora_alpha=16,
+        lora_dropout=0,
+        bias="none",
+        task_type="CAUSAL_LM",
+        ),
+    )
+    max_audio_samples = int(args.max_audio_seconds * 24000) + 1
+    train_rows, train_failures = preprocess(ds, train_idx, processor, max_audio_samples)
+    eval_rows, eval_failures = preprocess(ds, eval_idx, processor, max_audio_samples)
+
+    base_audio = generate_arm(model, processor, "base", args.output, args.generation_tokens)
+    trainer = Trainer(
+        model=model,
+        train_dataset=train_rows,
+        eval_dataset=eval_rows,
+        data_collator=TensorCollator(),
+        args=TrainingArguments(
+            output_dir=str(args.output / "sesame_checkpoints"),
+            per_device_train_batch_size=1,
+            per_device_eval_batch_size=1,
+            gradient_accumulation_steps=4,
+            max_steps=args.steps,
+            warmup_steps=5,
+            learning_rate=2e-4,
+            bf16=False,
+            logging_steps=1,
+            eval_strategy="no",
+            save_strategy="no",
+            optim="adamw_8bit",
+            weight_decay=0.01,
+            lr_scheduler_type="linear",
+            seed=SEED,
+            report_to="none",
+            remove_unused_columns=False,
+        ),
+    )
+    pre_eval = trainer.evaluate()
+    train_result = trainer.train()
+    post_eval = trainer.evaluate()
+    adapter_dir = args.output / "adapters" / "sesame"
+    # The temporary tokenizer snapshot contains only metadata/symlinks. Ensure
+    # the serialized adapter points consumers at the real public base model.
+    model.peft_config["default"].base_model_name_or_path = BASE_MODEL
+    model.save_pretrained(adapter_dir)
+    processor.save_pretrained(adapter_dir)
+    model.push_to_hub(args.hf_repo, private=False, token=os.environ.get("HF_TOKEN"))
+    processor.push_to_hub(args.hf_repo, private=False, token=os.environ.get("HF_TOKEN"))
+    adapter_revision = HfApi().model_info(args.hf_repo).sha
+    adapted_audio = generate_arm(model, processor, "adapted", args.output, args.generation_tokens)
+
+    adapter_files = [
+        {"path": str(x.relative_to(args.output)), "bytes": x.stat().st_size, "sha256": sha256(x)}
+        for x in sorted(adapter_dir.rglob("*"))
+        if x.is_file()
+    ]
+    manifest = {
+        "experiment": "8-6",
+        "track": "sesame_paralinguistic_tags",
+        "status": "trained_and_generated",
+        "seed": SEED,
+        "base_model": BASE_MODEL,
+        "base_model_revision": HfApi().model_info(BASE_MODEL).sha,
+        "dataset": DATASET,
+        "dataset_revision": DATASET_REVISION,
+        "source_dataset_note": "Public non-disabled mirror of the disabled MrDragonFox/Elise dataset named by the upstream notebook.",
+        "compatibility_notes": [
+            "Unsloth 2026.8.2 rejects CSM's tokenizer because the upstream tokenizer aliases pad to EOS; the run uses standard PEFT and CSM config token 128004 (<|finetune_right_pad_id|>).",
+            "Transformers 4.57 CSM codec embeddings are float32; float32 model/training avoids the bf16 merge dtype mismatch observed during the first pre-training evaluation attempt.",
+        ],
+        "train_examples_selected": len(train_idx),
+        "train_examples_preprocessed": len(train_rows),
+        "train_category_counts": train_counts,
+        "eval_examples_selected": len(eval_idx),
+        "eval_examples_preprocessed": len(eval_rows),
+        "eval_category_counts": eval_counts,
+        "train_failures": train_failures,
+        "eval_failures": eval_failures,
+        "max_audio_seconds": args.max_audio_seconds,
+        "optimizer_steps": args.steps,
+        "effective_batch_size": 4,
+        "lora_rank": 16,
+        "pre_eval": pre_eval,
+        "train_metrics": train_result.metrics,
+        "post_eval": post_eval,
+        "gpu": torch.cuda.get_device_name(0),
+        "peak_gpu_memory_bytes": torch.cuda.max_memory_reserved(),
+        "wall_seconds": time.time() - started,
+        "adapter_local_files": adapter_files,
+        "adapter_huggingface_repo": f"https://huggingface.co/{args.hf_repo}",
+        "adapter_huggingface_revision": adapter_revision,
+        "audio": base_audio + adapted_audio,
+    }
+    (args.output / "sesame_manifest.json").write_text(json.dumps(manifest, indent=2) + "\n", encoding="utf-8")
+
+
+if __name__ == "__main__":
+    main()

chapter8/speech-sft-experiment/test_experiment.py

@@ -0,0 +1,49 @@
+import importlib.util
+from pathlib import Path
+
+import torch
+
+
+HERE = Path(__file__).parent
+
+
+def load(name):
+    spec = importlib.util.spec_from_file_location(name, HERE / f"{name}.py")
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+    return module
+
+
+def test_sesame_tag_categories_are_explicit():
+    sesame = load("run_sesame")
+    assert sesame.category("hello <laughs> there") == "laugh"
+    assert sesame.category("hello <giggle> there") == "giggle"
+    assert sesame.category("hello <sighs> there") == "sigh"
+    assert sesame.category("hello there") == "neutral"
+
+
+def test_orpheus_collator_masks_label_padding():
+    orpheus = load("run_orpheus")
+    rows = [
+        {"input_ids": [1, 2], "labels": [1, 2], "attention_mask": [1, 1]},
+        {"input_ids": [3], "labels": [3], "attention_mask": [1]},
+    ]
+    batch = orpheus.PadCollator(9)(rows)
+    assert batch["input_ids"].tolist() == [[1, 2], [3, 9]]
+    assert batch["labels"].tolist() == [[1, 2], [3, -100]]
+    assert batch["attention_mask"].tolist() == [[1, 1], [1, 0]]
+
+
+def test_sesame_collator_stacks_all_model_inputs():
+    sesame = load("run_sesame")
+    rows = [{"input_ids": torch.tensor([1, 2]), "labels": torch.tensor([3, 4])}] * 2
+    batch = sesame.TensorCollator()(rows)
+    assert batch["input_ids"].shape == (2, 2)
+    assert batch["labels"].shape == (2, 2)
+
+
+def test_sha256_is_stable(tmp_path):
+    analysis = load("analyze_campaign")
+    path = tmp_path / "artifact"
+    path.write_bytes(b"experiment-8-6")
+    assert analysis.sha256(path) == "b07a691b33e493299473b6323258c9d643b2981d6de43e8c8adc3c4edc222d15"

chapter8/speech-sft-experiment/validation/exp8-6-20260804-v1/REPORT.md

@@ -0,0 +1,51 @@
+# Experiment 8-6 strict acceptance report
+
+Execution acceptance: **PASS**
+
+This run trained two real LoRA adapters on an RTX PRO 6000. It used 128 Orpheus training utterances plus 16 held-out utterances, and 168 stratified Sesame training utterances plus 24 held-out utterances. Each track completed 60 optimizer updates at effective batch size four. Both adapters are identified by local SHA-256 inventories and public Hugging Face repositories.
+
+## Execution gates
+
+- PASS — `orpheus_128_train_examples`
+- PASS — `orpheus_16_held_out_examples`
+- PASS — `orpheus_60_optimizer_steps`
+- PASS — `orpheus_remote_adapter_sha256_verified`
+- PASS — `orpheus_16_valid_comparison_files`
+- PASS — `sesame_128_train_examples`
+- PASS — `sesame_tag_categories_present`
+- PASS — `sesame_60_optimizer_steps`
+- PASS — `sesame_remote_adapter_sha256_verified`
+- PASS — `sesame_24_valid_comparison_files`
+
+## Hypothesis results
+
+- SUPPORTED — `orpheus_held_out_loss_decreased`
+- NOT SUPPORTED — `orpheus_cross_sentence_timbre_proxy_improved`
+- SUPPORTED — `sesame_held_out_loss_decreased`
+- SUPPORTED — `sesame_adapted_mean_tag_score_is_positive`
+- SUPPORTED — `sesame_tag_sensitivity_improved_over_base`
+
+Execution completion and hypothesis support are intentionally separate. A completed campaign may produce a negative hypothesis result.
+
+## Orpheus result
+
+- Held-out loss: 5.237792 before → 4.865821 after.
+- Mean cross-sentence MFCC-statistic cosine: 0.988627 base → 0.986702 adapted (Δ -0.001924).
+- Eight unseen sentences were generated for each arm with matched seeds. This metric is a timbre-consistency proxy; it is not speaker-verification or a listening-test score.
+
+## Sesame result
+
+- Held-out loss: 128.230759 before → 124.342400 after.
+- Mean matching AudioSet event-score difference (tagged − neutral): +0.000131 base → +0.001097 adapted (Δ +0.000966).
+- Positive matched pairs: 3/6 base; 4/6 adapted.
+- Six prompt pairs (laugh, giggle, sigh) were generated per arm with the same seed within each tagged/neutral pair. AudioSet scores are detector proxies, not proof of natural expression.
+
+## Failure retention and limits
+
+`failure_comparisons.json` retains silent/short outputs, each Orpheus arm's least-consistent sentence pair, and every Sesame pair where adding a tag did not raise the matching AudioSet score. `compatibility_failures.json` retains the disabled-source-dataset failure, current Unsloth CSM pad-token rejection, and Transformers bf16 codec merge failure, together with the exact standard-PEFT/float32 fallback. The Sesame held-out loss split contains laugh, sigh, and neutral examples but no giggle examples because all 32 available giggle-tagged rows were allocated to the substantive training split. The campaign does not include blinded human MOS, speaker-verification enrollment, confidence intervals over multiple training seeds, or deployment-scale data. Therefore it makes no claim of perceptual quality or generalization beyond this bounded run.
+
+## Adapter identity
+
+- Orpheus: https://huggingface.co/bojieli/exp8-6-orpheus-elise-lora/tree/536092e9479fa1717e2b8f9cc1be52728b273e95
+- Sesame: https://huggingface.co/bojieli/exp8-6-sesame-elise-tags-lora/tree/f2e042be0f38d6078976ef7e16cf49b91097f756
+- Exact revisions and every retained artifact hash are in `orpheus_manifest.json`, `sesame_manifest.json`, and `artifact_inventory.json`.

chapter8/speech-sft-experiment/validation/exp8-6-20260804-v1/adapters/orpheus/README.md

@@ -0,0 +1,208 @@
+---
+base_model: unsloth/orpheus-3b-0.1-ft
+library_name: peft
+pipeline_tag: text-generation
+tags:
+- base_model:adapter:unsloth/orpheus-3b-0.1-ft
+- lora
+- transformers
+- unsloth
+---
+
+# Model Card for Model ID
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+
+
+- **Developed by:** [More Information Needed]
+- **Funded by [optional]:** [More Information Needed]
+- **Shared by [optional]:** [More Information Needed]
+- **Model type:** [More Information Needed]
+- **Language(s) (NLP):** [More Information Needed]
+- **License:** [More Information Needed]
+- **Finetuned from model [optional]:** [More Information Needed]
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [More Information Needed]
+- **Paper [optional]:** [More Information Needed]
+- **Demo [optional]:** [More Information Needed]
+
+## Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+
+### Direct Use
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+[More Information Needed]
+
+### Downstream Use [optional]
+
+<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+
+[More Information Needed]
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+### Recommendations
+
+<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+
+Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+[More Information Needed]
+
+## Training Details
+
+### Training Data
+
+<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
+
+[More Information Needed]
+
+### Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+
+#### Preprocessing [optional]
+
+[More Information Needed]
+
+
+#### Training Hyperparameters
+
+- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+
+#### Speeds, Sizes, Times [optional]
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+[More Information Needed]
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+<!-- This should link to a Dataset Card if possible. -->
+
+[More Information Needed]
+
+#### Factors
+
+<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+
+[More Information Needed]
+
+#### Metrics
+
+<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+
+[More Information Needed]
+
+### Results
+
+[More Information Needed]
+
+#### Summary
+
+
+
+## Model Examination [optional]
+
+<!-- Relevant interpretability work for the model goes here -->
+
+[More Information Needed]
+
+## Environmental Impact
+
+<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+
+Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
+
+- **Hardware Type:** [More Information Needed]
+- **Hours used:** [More Information Needed]
+- **Cloud Provider:** [More Information Needed]
+- **Compute Region:** [More Information Needed]
+- **Carbon Emitted:** [More Information Needed]
+
+## Technical Specifications [optional]
+
+### Model Architecture and Objective
+
+[More Information Needed]
+
+### Compute Infrastructure
+
+[More Information Needed]
+
+#### Hardware
+
+[More Information Needed]
+
+#### Software
+
+[More Information Needed]
+
+## Citation [optional]
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+
+[More Information Needed]
+
+**APA:**
+
+[More Information Needed]
+
+## Glossary [optional]
+
+<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+
+[More Information Needed]
+
+## More Information [optional]
+
+[More Information Needed]
+
+## Model Card Authors [optional]
+
+[More Information Needed]
+
+## Model Card Contact
+
+[More Information Needed]
+### Framework versions
+
+- PEFT 0.19.0

chapter8/speech-sft-experiment/validation/exp8-6-20260804-v1/adapters/orpheus/adapter_config.json

@@ -0,0 +1,52 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": {
+    "base_model_class": "LlamaForCausalLM",
+    "parent_library": "transformers.models.llama.modeling_llama",
+    "unsloth_fixed": true
+  },
+  "base_model_name_or_path": "unsloth/orpheus-3b-0.1-ft",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 16,
+  "lora_bias": false,
+  "lora_dropout": 0,
+  "lora_ga_config": null,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "peft_version": "0.19.0",
+  "qalora_group_size": 16,
+  "r": 16,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "gate_proj",
+    "up_proj",
+    "o_proj",
+    "down_proj",
+    "k_proj",
+    "q_proj",
+    "v_proj"
+  ],
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_bdlora": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}

chapter8/speech-sft-experiment/validation/exp8-6-20260804-v1/adapters/orpheus/chat_template.jinja

@@ -0,0 +1,93 @@
+{{- bos_token }}
+{%- if custom_tools is defined %}
+    {%- set tools = custom_tools %}
+{%- endif %}
+{%- if not tools_in_user_message is defined %}
+    {%- set tools_in_user_message = true %}
+{%- endif %}
+{%- if not date_string is defined %}
+    {%- if strftime_now is defined %}
+        {%- set date_string = strftime_now("%d %b %Y") %}
+    {%- else %}
+        {%- set date_string = "26 Jul 2024" %}
+    {%- endif %}
+{%- endif %}
+{%- if not tools is defined %}
+    {%- set tools = none %}
+{%- endif %}
+
+{#- This block extracts the system message, so we can slot it into the right place. #}
+{%- if messages[0]['role'] == 'system' %}
+    {%- set system_message = messages[0]['content']|trim %}
+    {%- set messages = messages[1:] %}
+{%- else %}
+    {%- set system_message = "" %}
+{%- endif %}
+
+{#- System message #}
+{{- "<|start_header_id|>system<|end_header_id|>\n\n" }}
+{%- if tools is not none %}
+    {{- "Environment: ipython\n" }}
+{%- endif %}
+{{- "Cutting Knowledge Date: December 2023\n" }}
+{{- "Today Date: " + date_string + "\n\n" }}
+{%- if tools is not none and not tools_in_user_message %}
+    {{- "You have access to the following functions. To call a function, please respond with JSON for a function call." }}
+    {{- 'Respond in the format {"name": function name, "parameters": dictionary of argument name and its value}.' }}
+    {{- "Do not use variables.\n\n" }}
+    {%- for t in tools %}
+        {{- t | tojson(indent=4) }}
+        {{- "\n\n" }}
+    {%- endfor %}
+{%- endif %}
+{{- system_message }}
+{{- "<|eot_id|>" }}
+
+{#- Custom tools are passed in a user message with some extra guidance #}
+{%- if tools_in_user_message and not tools is none %}
+    {#- Extract the first user message so we can plug it in here #}
+    {%- if messages | length != 0 %}
+        {%- set first_user_message = messages[0]['content']|trim %}
+        {%- set messages = messages[1:] %}
+    {%- else %}
+        {{- raise_exception("Cannot put tools in the first user message when there's no first user message!") }}
+{%- endif %}
+    {{- '<|start_header_id|>user<|end_header_id|>\n\n' -}}
+    {{- "Given the following functions, please respond with a JSON for a function call " }}
+    {{- "with its proper arguments that best answers the given prompt.\n\n" }}
+    {{- 'Respond in the format {"name": function name, "parameters": dictionary of argument name and its value}.' }}
+    {{- "Do not use variables.\n\n" }}
+    {%- for t in tools %}
+        {{- t | tojson(indent=4) }}
+        {{- "\n\n" }}
+    {%- endfor %}
+    {{- first_user_message + "<|eot_id|>"}}
+{%- endif %}
+
+{%- for message in messages %}
+    {%- if not (message.role == 'ipython' or message.role == 'tool' or 'tool_calls' in message) %}
+        {{- '<|start_header_id|>' + message['role'] + '<|end_header_id|>\n\n'+ message['content'] | trim + '<|eot_id|>' }}
+    {%- elif 'tool_calls' in message %}
+        {%- if not message.tool_calls|length == 1 %}
+            {{- raise_exception("This model only supports single tool-calls at once!") }}
+        {%- endif %}
+        {%- set tool_call = message.tool_calls[0].function %}
+        {{- '<|start_header_id|>assistant<|end_header_id|>\n\n' -}}
+        {{- '{"name": "' + tool_call.name + '", ' }}
+        {{- '"parameters": ' }}
+        {{- tool_call.arguments | tojson }}
+        {{- "}" }}
+        {{- "<|eot_id|>" }}
+    {%- elif message.role == "tool" or message.role == "ipython" %}
+        {{- "<|start_header_id|>ipython<|end_header_id|>\n\n" }}
+        {%- if message.content is mapping or message.content is iterable %}
+            {{- message.content | tojson }}
+        {%- else %}
+            {{- message.content }}
+        {%- endif %}
+        {{- "<|eot_id|>" }}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|start_header_id|>assistant<|end_header_id|>\n\n' }}
+{%- endif %}

chapter8/speech-sft-experiment/validation/exp8-6-20260804-v1/adapters/orpheus/special_tokens_map.json

@@ -0,0 +1,26 @@
+{
+  "additional_special_tokens": [
+    "<|audio|>"
+  ],
+  "bos_token": {
+    "content": "<|begin_of_text|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "<|eot_id|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<|finetune_right_pad_id|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

chapter8/speech-sft-experiment/validation/exp8-6-20260804-v1/adapters/sesame/README.md

@@ -0,0 +1,207 @@
+---
+base_model: unsloth/csm-1b
+library_name: peft
+pipeline_tag: text-generation
+tags:
+- base_model:adapter:unsloth/csm-1b
+- lora
+- transformers
+---
+
+# Model Card for Model ID
+
+<!-- Provide a quick summary of what the model is/does. -->
+
+
+
+## Model Details
+
+### Model Description
+
+<!-- Provide a longer summary of what this model is. -->
+
+
+
+- **Developed by:** [More Information Needed]
+- **Funded by [optional]:** [More Information Needed]
+- **Shared by [optional]:** [More Information Needed]
+- **Model type:** [More Information Needed]
+- **Language(s) (NLP):** [More Information Needed]
+- **License:** [More Information Needed]
+- **Finetuned from model [optional]:** [More Information Needed]
+
+### Model Sources [optional]
+
+<!-- Provide the basic links for the model. -->
+
+- **Repository:** [More Information Needed]
+- **Paper [optional]:** [More Information Needed]
+- **Demo [optional]:** [More Information Needed]
+
+## Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+
+### Direct Use
+
+<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
+
+[More Information Needed]
+
+### Downstream Use [optional]
+
+<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
+
+[More Information Needed]
+
+### Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+
+[More Information Needed]
+
+## Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+
+[More Information Needed]
+
+### Recommendations
+
+<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
+
+Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
+
+## How to Get Started with the Model
+
+Use the code below to get started with the model.
+
+[More Information Needed]
+
+## Training Details
+
+### Training Data
+
+<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
+
+[More Information Needed]
+
+### Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+
+#### Preprocessing [optional]
+
+[More Information Needed]
+
+
+#### Training Hyperparameters
+
+- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+
+#### Speeds, Sizes, Times [optional]
+
+<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+
+[More Information Needed]
+
+## Evaluation
+
+<!-- This section describes the evaluation protocols and provides the results. -->
+
+### Testing Data, Factors & Metrics
+
+#### Testing Data
+
+<!-- This should link to a Dataset Card if possible. -->
+
+[More Information Needed]
+
+#### Factors
+
+<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+
+[More Information Needed]
+
+#### Metrics
+
+<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+
+[More Information Needed]
+
+### Results
+
+[More Information Needed]
+
+#### Summary
+
+
+
+## Model Examination [optional]
+
+<!-- Relevant interpretability work for the model goes here -->
+
+[More Information Needed]
+
+## Environmental Impact
+
+<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
+
+Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
+
+- **Hardware Type:** [More Information Needed]
+- **Hours used:** [More Information Needed]
+- **Cloud Provider:** [More Information Needed]
+- **Compute Region:** [More Information Needed]
+- **Carbon Emitted:** [More Information Needed]
+
+## Technical Specifications [optional]
+
+### Model Architecture and Objective
+
+[More Information Needed]
+
+### Compute Infrastructure
+
+[More Information Needed]
+
+#### Hardware
+
+[More Information Needed]
+
+#### Software
+
+[More Information Needed]
+
+## Citation [optional]
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+
+[More Information Needed]
+
+**APA:**
+
+[More Information Needed]
+
+## Glossary [optional]
+
+<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
+
+[More Information Needed]
+
+## More Information [optional]
+
+[More Information Needed]
+
+## Model Card Authors [optional]
+
+[More Information Needed]
+
+## Model Card Contact
+
+[More Information Needed]
+### Framework versions
+
+- PEFT 0.19.0

chapter8/speech-sft-experiment/validation/exp8-6-20260804-v1/adapters/sesame/adapter_config.json

@@ -0,0 +1,48 @@
+{
+  "alora_invocation_tokens": null,
+  "alpha_pattern": {},
+  "arrow_config": null,
+  "auto_mapping": null,
+  "base_model_name_or_path": "unsloth/csm-1b",
+  "bias": "none",
+  "corda_config": null,
+  "ensure_weight_tying": false,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 16,
+  "lora_bias": false,
+  "lora_dropout": 0,
+  "lora_ga_config": null,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "peft_version": "0.19.0",
+  "qalora_group_size": 16,
+  "r": 16,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "down_proj",
+    "v_proj",
+    "q_proj",
+    "gate_proj",
+    "k_proj",
+    "o_proj",
+    "up_proj"
+  ],
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_bdlora": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}

chapter8/speech-sft-experiment/validation/exp8-6-20260804-v1/adapters/sesame/chat_template.jinja

@@ -0,0 +1,45 @@
+
+{%- for message in messages %}
+    {#-- Validate role is a stringified integer --#}
+    {%- if not message['role'] is string or not message['role'].isdigit() %}
+        {{- raise_exception("The role must be an integer or a stringified integer (e.g. '0') designating the speaker id") }}
+    {%- endif %}
+
+    {#-- Validate content is a list --#}
+    {%- set content = message['content'] %}
+    {%- if content is not iterable or content is string %}
+        {{- raise_exception("The content must be a list") }}
+    {%- endif %}
+
+    {#-- Collect content types --#}
+    {%- set content_types = content | map(attribute='type') | list %}
+    {%- set is_last = loop.last %}
+
+    {#-- Last message validation --#}
+    {%- if is_last %}
+        {%- if 'text' not in content_types %}
+            {{- raise_exception("The last message must include one item of type 'text'") }}
+        {%- elif (content_types | select('equalto', 'text') | list | length > 1) or (content_types | select('equalto', 'audio') | list | length > 1) %}
+            {{- raise_exception("At most two items are allowed in the last message: one 'text' and one 'audio'") }}
+        {%- endif %}
+
+    {#-- All other messages validation --#}
+    {%- else %}
+        {%- if content_types | select('equalto', 'text') | list | length != 1
+              or content_types | select('equalto', 'audio') | list | length != 1 %}
+            {{- raise_exception("Each message (except the last) must contain exactly one 'text' and one 'audio' item") }}
+        {%- elif content_types | reject('in', ['text', 'audio']) | list | length > 0 %}
+            {{- raise_exception("Only 'text' and 'audio' types are allowed in content") }}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+
+{%- for message in messages %}
+    {{- bos_token }}
+    {{- '[' + message['role'] + ']' }}
+    {{- message['content'][0]['text'] }}
+    {{- eos_token }}
+    {%- if message['content']|length > 1 %}
+        {{- '<|AUDIO|><|audio_eos|>' }}
+    {%- endif %}
+{%- endfor %}

chapter8/speech-sft-experiment/validation/exp8-6-20260804-v1/adapters/sesame/preprocessor_config.json

@@ -0,0 +1,11 @@
+{
+  "chunk_length_s": null,
+  "feature_extractor_type": "EncodecFeatureExtractor",
+  "feature_size": 1,
+  "overlap": null,
+  "padding_side": "right",
+  "padding_value": 0.0,
+  "processor_class": "CsmProcessor",
+  "return_attention_mask": true,
+  "sampling_rate": 24000
+}

chapter8/speech-sft-experiment/validation/exp8-6-20260804-v1/adapters/sesame/special_tokens_map.json

@@ -0,0 +1,23 @@
+{
+  "bos_token": {
+    "content": "<|begin_of_text|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "<|end_of_text|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<|finetune_right_pad_id|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

chapter8/speech-sft-experiment/validation/exp8-6-20260804-v1/analysis.json

@@ -0,0 +1,723 @@
+{
+  "experiment": "8-6",
+  "execution_acceptance": "PASS",
+  "execution_gates": {
+    "orpheus_128_train_examples": true,
+    "orpheus_16_held_out_examples": true,
+    "orpheus_60_optimizer_steps": true,
+    "orpheus_remote_adapter_sha256_verified": true,
+    "orpheus_16_valid_comparison_files": true,
+    "sesame_128_train_examples": true,
+    "sesame_tag_categories_present": true,
+    "sesame_60_optimizer_steps": true,
+    "sesame_remote_adapter_sha256_verified": true,
+    "sesame_24_valid_comparison_files": true
+  },
+  "hypothesis_results": {
+    "orpheus_held_out_loss_decreased": true,
+    "orpheus_cross_sentence_timbre_proxy_improved": false,
+    "sesame_held_out_loss_decreased": true,
+    "sesame_adapted_mean_tag_score_is_positive": true,
+    "sesame_tag_sensitivity_improved_over_base": true
+  },
+  "quality_claim": "No human naturalness or voice-identity quality claim; automatic metrics are reproducible proxies only.",
+  "remote_adapter_verification": {
+    "orpheus": {
+      "repository": "https://huggingface.co/bojieli/exp8-6-orpheus-elise-lora",
+      "revision": "536092e9479fa1717e2b8f9cc1be52728b273e95",
+      "expected_sha256": "fbd43cd5287a4b69398c29f88286e60146e9328cb48e3c22e15a2a305b25b29f",
+      "downloaded_sha256": "fbd43cd5287a4b69398c29f88286e60146e9328cb48e3c22e15a2a305b25b29f",
+      "verified": true
+    },
+    "sesame": {
+      "repository": "https://huggingface.co/bojieli/exp8-6-sesame-elise-tags-lora",
+      "revision": "f2e042be0f38d6078976ef7e16cf49b91097f756",
+      "expected_sha256": "23994b467b319f99bbc642f4dc2ba4d39718838cd7cbfe949311a42ebcb5480e",
+      "downloaded_sha256": "23994b467b319f99bbc642f4dc2ba4d39718838cd7cbfe949311a42ebcb5480e",
+      "verified": true
+    }
+  },
+  "orpheus": {
+    "base": {
+      "audio_count": 8,
+      "valid_audio_count": 8,
+      "mean_pairwise_mfcc_cosine": 0.9886267249073301,
+      "min_pairwise_mfcc_cosine": 0.9738007187843323,
+      "pairwise": [
+        {
+          "prompt_a": 1,
+          "prompt_b": 7,
+          "cosine": 0.9738007187843323
+        },
+        {
+          "prompt_a": 3,
+          "prompt_b": 7,
+          "cosine": 0.9751747846603394
+        },
+        {
+          "prompt_a": 5,
+          "prompt_b": 7,
+          "cosine": 0.9768423438072205
+        },
+        {
+          "prompt_a": 1,
+          "prompt_b": 6,
+          "cosine": 0.9821145534515381
+        },
+        {
+          "prompt_a": 5,
+          "prompt_b": 6,
+          "cosine": 0.9832608699798584
+        },
+        {
+          "prompt_a": 0,
+          "prompt_b": 7,
+          "cosine": 0.9844203591346741
+        },
+        {
+          "prompt_a": 3,
+          "prompt_b": 4,
+          "cosine": 0.9847696423530579
+        },
+        {
+          "prompt_a": 4,
+          "prompt_b": 5,
+          "cosine": 0.985227644443512
+        },
+        {
+          "prompt_a": 2,
+          "prompt_b": 7,
+          "cosine": 0.9852697849273682
+        },
+        {
+          "prompt_a": 1,
+          "prompt_b": 4,
+          "cosine": 0.9852963089942932
+        },
+        {
+          "prompt_a": 3,
+          "prompt_b": 6,
+          "cosine": 0.9860416650772095
+        },
+        {
+          "prompt_a": 0,
+          "prompt_b": 6,
+          "cosine": 0.9888354539871216
+        },
+        {
+          "prompt_a": 0,
+          "prompt_b": 1,
+          "cosine": 0.9906319379806519
+        },
+        {
+          "prompt_a": 0,
+          "prompt_b": 5,
+          "cosine": 0.9906508326530457
+        },
+        {
+          "prompt_a": 0,
+          "prompt_b": 4,
+          "cosine": 0.991395890712738
+        },
+        {
+          "prompt_a": 3,
+          "prompt_b": 5,
+          "cosine": 0.9918218851089478
+        },
+        {
+          "prompt_a": 1,
+          "prompt_b": 3,
+          "cosine": 0.9923202395439148
+        },
+        {
+          "prompt_a": 1,
+          "prompt_b": 2,
+          "cosine": 0.9923385977745056
+        },
+        {
+          "prompt_a": 2,
+          "prompt_b": 3,
+          "cosine": 0.9924396276473999
+        },
+        {
+          "prompt_a": 4,
+          "prompt_b": 7,
+          "cosine": 0.9930636882781982
+        },
+        {
+          "prompt_a": 2,
+          "prompt_b": 4,
+          "cosine": 0.9931535124778748
+        },
+        {
+          "prompt_a": 2,
+          "prompt_b": 5,
+          "cosine": 0.9931651949882507
+        },
+        {
+          "prompt_a": 6,
+          "prompt_b": 7,
+          "cosine": 0.9940599799156189
+        },
+        {
+          "prompt_a": 2,
+          "prompt_b": 6,
+          "cosine": 0.9941393136978149
+        },
+        {
+          "prompt_a": 0,
+          "prompt_b": 3,
+          "cosine": 0.9943145513534546
+        },
+        {
+          "prompt_a": 0,
+          "prompt_b": 2,
+          "cosine": 0.994386613368988
+        },
+        {
+          "prompt_a": 4,
+          "prompt_b": 6,
+          "cosine": 0.9950653314590454
+        },
+        {
+          "prompt_a": 1,
+          "prompt_b": 5,
+          "cosine": 0.9975469708442688
+        }
+      ]
+    },
+    "adapted": {
+      "audio_count": 8,
+      "valid_audio_count": 8,
+      "mean_pairwise_mfcc_cosine": 0.9867024081093925,
+      "min_pairwise_mfcc_cosine": 0.9688550233840942,
+      "pairwise": [
+        {
+          "prompt_a": 1,
+          "prompt_b": 5,
+          "cosine": 0.9688550233840942
+        },
+        {
+          "prompt_a": 0,
+          "prompt_b": 5,
+          "cosine": 0.9705199003219604
+        },
+        {
+          "prompt_a": 3,
+          "prompt_b": 5,
+          "cosine": 0.9741944670677185
+        },
+        {
+          "prompt_a": 5,
+          "prompt_b": 6,
+          "cosine": 0.9782752990722656
+        },
+        {
+          "prompt_a": 4,
+          "prompt_b": 5,
+          "cosine": 0.9795999526977539
+        },
+        {
+          "prompt_a": 5,
+          "prompt_b": 7,
+          "cosine": 0.9798238277435303
+        },
+        {
+          "prompt_a": 0,
+          "prompt_b": 7,
+          "cosine": 0.9805128574371338
+        },
+        {
+          "prompt_a": 1,
+          "prompt_b": 2,
+          "cosine": 0.9849662184715271
+        },
+        {
+          "prompt_a": 0,
+          "prompt_b": 6,
+          "cosine": 0.9854229688644409
+        },
+        {
+          "prompt_a": 0,
+          "prompt_b": 2,
+          "cosine": 0.9863125681877136
+        },
+        {
+          "prompt_a": 0,
+          "prompt_b": 4,
+          "cosine": 0.9869258999824524
+        },
+        {
+          "prompt_a": 1,
+          "prompt_b": 4,
+          "cosine": 0.9872470498085022
+        },
+        {
+          "prompt_a": 2,
+          "prompt_b": 3,
+          "cosine": 0.9892308115959167
+        },
+        {
+          "prompt_a": 0,
+          "prompt_b": 1,
+          "cosine": 0.9895753860473633
+        },
+        {
+          "prompt_a": 2,
+          "prompt_b": 7,
+          "cosine": 0.9899733662605286
+        },
+        {
+          "prompt_a": 3,
+          "prompt_b": 6,
+          "cosine": 0.9900691509246826
+        },
+        {
+          "prompt_a": 0,
+          "prompt_b": 3,
+          "cosine": 0.9903662800788879
+        },
+        {
+          "prompt_a": 1,
+          "prompt_b": 3,
+          "cosine": 0.9910094141960144
+        },
+        {
+          "prompt_a": 3,
+          "prompt_b": 4,
+          "cosine": 0.9911966323852539
+        },
+        {
+          "prompt_a": 2,
+          "prompt_b": 6,
+          "cosine": 0.9914029836654663
+        },
+        {
+          "prompt_a": 2,
+          "prompt_b": 5,
+          "cosine": 0.9914451241493225
+        },
+        {
+          "prompt_a": 1,
+          "prompt_b": 7,
+          "cosine": 0.9918804168701172
+        },
+        {
+          "prompt_a": 4,
+          "prompt_b": 6,
+          "cosine": 0.9921923279762268
+        },
+        {
+          "prompt_a": 2,
+          "prompt_b": 4,
+          "cosine": 0.9927540421485901
+        },
+        {
+          "prompt_a": 1,
+          "prompt_b": 6,
+          "cosine": 0.992949903011322
+        },
+        {
+          "prompt_a": 3,
+          "prompt_b": 7,
+          "cosine": 0.993278980255127
+        },
+        {
+          "prompt_a": 6,
+          "prompt_b": 7,
+          "cosine": 0.9935612678527832
+        },
+        {
+          "prompt_a": 4,
+          "prompt_b": 7,
+          "cosine": 0.9941253066062927
+        }
+      ]
+    },
+    "adapted_minus_base_mean_pairwise_mfcc_cosine": -0.0019243167979375864
+  },
+  "sesame": {
+    "classifier": "MIT/ast-finetuned-audioset-10-10-0.4593",
+    "labels": {
+      "laugh": {
+        "id": 16,
+        "name": "Laughter"
+      },
+      "giggle": {
+        "id": 18,
+        "name": "Giggle"
+      },
+      "sigh": {
+        "id": 26,
+        "name": "Sigh"
+      }
+    },
+    "base": {
+      "audio_count": 12,
+      "valid_audio_count": 12,
+      "mean_tagged_minus_neutral": 0.00013129252571767816,
+      "positive_pair_count": 3,
+      "pairs": [
+        {
+          "pair_id": "giggle_0",
+          "tag": "giggle",
+          "neutral_score": 0.0009356054943054914,
+          "tagged_score": 0.0006116584991104901,
+          "tagged_minus_neutral": -0.00032394699519500136
+        },
+        {
+          "pair_id": "giggle_1",
+          "tag": "giggle",
+          "neutral_score": 0.0008636609418317676,
+          "tagged_score": 0.0005855335621163249,
+          "tagged_minus_neutral": -0.00027812737971544266
+        },
+        {
+          "pair_id": "laugh_0",
+          "tag": "laugh",
+          "neutral_score": 0.0003021236334461719,
+          "tagged_score": 0.0003340380499139428,
+          "tagged_minus_neutral": 3.1914416467770934e-05
+        },
+        {
+          "pair_id": "laugh_1",
+          "tag": "laugh",
+          "neutral_score": 0.0002934956573881209,
+          "tagged_score": 0.0007533965981565416,
+          "tagged_minus_neutral": 0.0004599009407684207
+        },
+        {
+          "pair_id": "sigh_0",
+          "tag": "sigh",
+          "neutral_score": 0.0034571753349155188,
+          "tagged_score": 0.006142919883131981,
+          "tagged_minus_neutral": 0.002685744548216462
+        },
+        {
+          "pair_id": "sigh_1",
+          "tag": "sigh",
+          "neutral_score": 0.0024482603184878826,
+          "tagged_score": 0.0006605299422517419,
+          "tagged_minus_neutral": -0.0017877303762361407
+        }
+      ]
+    },
+    "adapted": {
+      "audio_count": 12,
+      "valid_audio_count": 12,
+      "mean_tagged_minus_neutral": 0.001097210906057929,
+      "positive_pair_count": 4,
+      "pairs": [
+        {
+          "pair_id": "giggle_0",
+          "tag": "giggle",
+          "neutral_score": 0.001688887132331729,
+          "tagged_score": 0.0017779936315491796,
+          "tagged_minus_neutral": 8.910649921745062e-05
+        },
+        {
+          "pair_id": "giggle_1",
+          "tag": "giggle",
+          "neutral_score": 0.003269059816375375,
+          "tagged_score": 0.0007443267968483269,
+          "tagged_minus_neutral": -0.002524733019527048
+        },
+        {
+          "pair_id": "laugh_0",
+          "tag": "laugh",
+          "neutral_score": 0.00026441123918630183,
+          "tagged_score": 0.0012162356870248914,
+          "tagged_minus_neutral": 0.0009518244478385895
+        },
+        {
+          "pair_id": "laugh_1",
+          "tag": "laugh",
+          "neutral_score": 0.00033708245609886944,
+          "tagged_score": 0.0003236289485357702,
+          "tagged_minus_neutral": -1.3453507563099265e-05
+        },
+        {
+          "pair_id": "sigh_0",
+          "tag": "sigh",
+          "neutral_score": 0.0017972267232835293,
+          "tagged_score": 0.006373145151883364,
+          "tagged_minus_neutral": 0.0045759184285998344
+        },
+        {
+          "pair_id": "sigh_1",
+          "tag": "sigh",
+          "neutral_score": 0.0010693982476368546,
+          "tagged_score": 0.004574000835418701,
+          "tagged_minus_neutral": 0.0035046025877818465
+        }
+      ]
+    },
+    "adapted_minus_base_mean_tag_sensitivity": 0.0009659183803402508,
+    "scores": [
+      {
+        "arm": "base",
+        "pair_id": "laugh_0",
+        "condition": "neutral",
+        "tag": "laugh",
+        "audioset_label": "Laughter",
+        "audioset_score": 0.0003021236334461719,
+        "seconds": 2.72,
+        "rms": 0.06748455762863159,
+        "path": "audio/sesame/base/laugh_0_neutral.wav"
+      },
+      {
+        "arm": "base",
+        "pair_id": "laugh_0",
+        "condition": "tagged",
+        "tag": "laugh",
+        "audioset_label": "Laughter",
+        "audioset_score": 0.0003340380499139428,
+        "seconds": 6.0,
+        "rms": 0.06531786173582077,
+        "path": "audio/sesame/base/laugh_0_tagged.wav"
+      },
+      {
+        "arm": "base",
+        "pair_id": "laugh_1",
+        "condition": "neutral",
+        "tag": "laugh",
+        "audioset_label": "Laughter",
+        "audioset_score": 0.0002934956573881209,
+        "seconds": 2.08,
+        "rms": 0.23473648726940155,
+        "path": "audio/sesame/base/laugh_1_neutral.wav"
+      },
+      {
+        "arm": "base",
+        "pair_id": "laugh_1",
+        "condition": "tagged",
+        "tag": "laugh",
+        "audioset_label": "Laughter",
+        "audioset_score": 0.0007533965981565416,
+        "seconds": 3.84,
+        "rms": 0.03746151924133301,
+        "path": "audio/sesame/base/laugh_1_tagged.wav"
+      },
+      {
+        "arm": "base",
+        "pair_id": "giggle_0",
+        "condition": "neutral",
+        "tag": "giggle",
+        "audioset_label": "Giggle",
+        "audioset_score": 0.0009356054943054914,
+        "seconds": 1.68,
+        "rms": 0.14035280048847198,
+        "path": "audio/sesame/base/giggle_0_neutral.wav"
+      },
+      {
+        "arm": "base",
+        "pair_id": "giggle_0",
+        "condition": "tagged",
+        "tag": "giggle",
+        "audioset_label": "Giggle",
+        "audioset_score": 0.0006116584991104901,
+        "seconds": 2.96,
+        "rms": 0.08793672919273376,
+        "path": "audio/sesame/base/giggle_0_tagged.wav"
+      },
+      {
+        "arm": "base",
+        "pair_id": "giggle_1",
+        "condition": "neutral",
+        "tag": "giggle",
+        "audioset_label": "Giggle",
+        "audioset_score": 0.0008636609418317676,
+        "seconds": 2.96,
+        "rms": 0.049566783010959625,
+        "path": "audio/sesame/base/giggle_1_neutral.wav"
+      },
+      {
+        "arm": "base",
+        "pair_id": "giggle_1",
+        "condition": "tagged",
+        "tag": "giggle",
+        "audioset_label": "Giggle",
+        "audioset_score": 0.0005855335621163249,
+        "seconds": 5.92,
+        "rms": 0.0142013905569911,
+        "path": "audio/sesame/base/giggle_1_tagged.wav"
+      },
+      {
+        "arm": "base",
+        "pair_id": "sigh_0",
+        "condition": "neutral",
+        "tag": "sigh",
+        "audioset_label": "Sigh",
+        "audioset_score": 0.0034571753349155188,
+        "seconds": 2.32,
+        "rms": 0.06743967533111572,
+        "path": "audio/sesame/base/sigh_0_neutral.wav"
+      },
+      {
+        "arm": "base",
+        "pair_id": "sigh_0",
+        "condition": "tagged",
+        "tag": "sigh",
+        "audioset_label": "Sigh",
+        "audioset_score": 0.006142919883131981,
+        "seconds": 6.0,
+        "rms": 0.05658983066678047,
+        "path": "audio/sesame/base/sigh_0_tagged.wav"
+      },
+      {
+        "arm": "base",
+        "pair_id": "sigh_1",
+        "condition": "neutral",
+        "tag": "sigh",
+        "audioset_label": "Sigh",
+        "audioset_score": 0.0024482603184878826,
+        "seconds": 2.24,
+        "rms": 0.014161595143377781,
+        "path": "audio/sesame/base/sigh_1_neutral.wav"
+      },
+      {
+        "arm": "base",
+        "pair_id": "sigh_1",
+        "condition": "tagged",
+        "tag": "sigh",
+        "audioset_label": "Sigh",
+        "audioset_score": 0.0006605299422517419,
+        "seconds": 3.36,
+        "rms": 0.010387708432972431,
+        "path": "audio/sesame/base/sigh_1_tagged.wav"
+      },
+      {
+        "arm": "adapted",
+        "pair_id": "laugh_0",
+        "condition": "neutral",
+        "tag": "laugh",
+        "audioset_label": "Laughter",
+        "audioset_score": 0.00026441123918630183,
+        "seconds": 3.92,
+        "rms": 0.0402710996568203,
+        "path": "audio/sesame/adapted/laugh_0_neutral.wav"
+      },
+      {
+        "arm": "adapted",
+        "pair_id": "laugh_0",
+        "condition": "tagged",
+        "tag": "laugh",
+        "audioset_label": "Laughter",
+        "audioset_score": 0.0012162356870248914,
+        "seconds": 3.6,
+        "rms": 0.06890808790922165,
+        "path": "audio/sesame/adapted/laugh_0_tagged.wav"
+      },
+      {
+        "arm": "adapted",
+        "pair_id": "laugh_1",
+        "condition": "neutral",
+        "tag": "laugh",
+        "audioset_label": "Laughter",
+        "audioset_score": 0.00033708245609886944,
+        "seconds": 2.48,
+        "rms": 0.13957583904266357,
+        "path": "audio/sesame/adapted/laugh_1_neutral.wav"
+      },
+      {
+        "arm": "adapted",
+        "pair_id": "laugh_1",
+        "condition": "tagged",
+        "tag": "laugh",
+        "audioset_label": "Laughter",
+        "audioset_score": 0.0003236289485357702,
+        "seconds": 3.76,
+        "rms": 0.051783934235572815,
+        "path": "audio/sesame/adapted/laugh_1_tagged.wav"
+      },
+      {
+        "arm": "adapted",
+        "pair_id": "giggle_0",
+        "condition": "neutral",
+        "tag": "giggle",
+        "audioset_label": "Giggle",
+        "audioset_score": 0.001688887132331729,
+        "seconds": 2.0,
+        "rms": 0.11314665526151657,
+        "path": "audio/sesame/adapted/giggle_0_neutral.wav"
+      },
+      {
+        "arm": "adapted",
+        "pair_id": "giggle_0",
+        "condition": "tagged",
+        "tag": "giggle",
+        "audioset_label": "Giggle",
+        "audioset_score": 0.0017779936315491796,
+        "seconds": 2.72,
+        "rms": 0.07838206738233566,
+        "path": "audio/sesame/adapted/giggle_0_tagged.wav"
+      },
+      {
+        "arm": "adapted",
+        "pair_id": "giggle_1",
+        "condition": "neutral",
+        "tag": "giggle",
+        "audioset_label": "Giggle",
+        "audioset_score": 0.003269059816375375,
+        "seconds": 3.28,
+        "rms": 0.10014152526855469,
+        "path": "audio/sesame/adapted/giggle_1_neutral.wav"
+      },
+      {
+        "arm": "adapted",
+        "pair_id": "giggle_1",
+        "condition": "tagged",
+        "tag": "giggle",
+        "audioset_label": "Giggle",
+        "audioset_score": 0.0007443267968483269,
+        "seconds": 4.88,
+        "rms": 0.07103203237056732,
+        "path": "audio/sesame/adapted/giggle_1_tagged.wav"
+      },
+      {
+        "arm": "adapted",
+        "pair_id": "sigh_0",
+        "condition": "neutral",
+        "tag": "sigh",
+        "audioset_label": "Sigh",
+        "audioset_score": 0.0017972267232835293,
+        "seconds": 2.72,
+        "rms": 0.09723222255706787,
+        "path": "audio/sesame/adapted/sigh_0_neutral.wav"
+      },
+      {
+        "arm": "adapted",
+        "pair_id": "sigh_0",
+        "condition": "tagged",
+        "tag": "sigh",
+        "audioset_label": "Sigh",
+        "audioset_score": 0.006373145151883364,
+        "seconds": 3.92,
+        "rms": 0.1117464229464531,
+        "path": "audio/sesame/adapted/sigh_0_tagged.wav"
+      },
+      {
+        "arm": "adapted",
+        "pair_id": "sigh_1",
+        "condition": "neutral",
+        "tag": "sigh",
+        "audioset_label": "Sigh",
+        "audioset_score": 0.0010693982476368546,
+        "seconds": 2.32,
+        "rms": 0.15041208267211914,
+        "path": "audio/sesame/adapted/sigh_1_neutral.wav"
+      },
+      {
+        "arm": "adapted",
+        "pair_id": "sigh_1",
+        "condition": "tagged",
+        "tag": "sigh",
+        "audioset_label": "Sigh",
+        "audioset_score": 0.004574000835418701,
+        "seconds": 3.36,
+        "rms": 0.07697359472513199,
+        "path": "audio/sesame/adapted/sigh_1_tagged.wav"
+      }
+    ]
+  }
+}

chapter8/speech-sft-experiment/validation/exp8-6-20260804-v1/artifact_inventory.json

@@ -0,0 +1,320 @@
+[
+  {
+    "path": "adapters/orpheus/README.md",
+    "bytes": 5214,
+    "sha256": "d62b94e4ab068c9379c9d6a92fa73dc9468f85ce204615543fa20db6e453e4fb"
+  },
+  {
+    "path": "adapters/orpheus/adapter_config.json",
+    "bytes": 1238,
+    "sha256": "40bd5beba0cbeb72bbe0acf0371d0ab64f1217722d3590ee165f5f9d2f745ca9"
+  },
+  {
+    "path": "adapters/orpheus/chat_template.jinja",
+    "bytes": 3827,
+    "sha256": "5816fce10444e03c2e9ee1ef8a4a1ea61ae7e69e438613f3b17b69d0426223a4"
+  },
+  {
+    "path": "adapters/orpheus/special_tokens_map.json",
+    "bytes": 508,
+    "sha256": "f17390e64274b3117a7f99afa87eb1de93097b820842c68bd626aff53d73d7f5"
+  },
+  {
+    "path": "adapters/sesame/README.md",
+    "bytes": 5183,
+    "sha256": "65a8c53feecbfc1e7a8cd64bc6a79baedbffd1ef6b9cc17b1ea0a5529e4d74ec"
+  },
+  {
+    "path": "adapters/sesame/adapter_config.json",
+    "bytes": 1091,
+    "sha256": "44c3053b5037214ea8862003dfeea867554c9fbb83df5728ea80e9a0915e3ee2"
+  },
+  {
+    "path": "adapters/sesame/chat_template.jinja",
+    "bytes": 2002,
+    "sha256": "b43753fbbc23be93160cccf825d2cf97dec39538cb274350216940e7acbc0fe7"
+  },
+  {
+    "path": "adapters/sesame/preprocessor_config.json",
+    "bytes": 271,
+    "sha256": "8648040cda0469976874c733acbec88c0ef094cd95628ab63b32950f01872af9"
+  },
+  {
+    "path": "adapters/sesame/special_tokens_map.json",
+    "bytes": 459,
+    "sha256": "32f404d626cf7b1b6eea36c241ae7cbd6ec29c777a8701ea48c0fe9ebb94c9b1"
+  },
+  {
+    "path": "analysis.json",
+    "bytes": 19921,
+    "sha256": "958f7bb5188233867f50ee4920a4bc8d4ae46b1af1eb46ca8eef74cde88e3f1c"
+  },
+  {
+    "path": "audio/orpheus/adapted/prompt_00.wav",
+    "bytes": 221228,
+    "sha256": "24470415a507117b9f2b6a4b3f13350a453febca782f542e592ea6d41678f9da"
+  },
+  {
+    "path": "audio/orpheus/adapted/prompt_01.wav",
+    "bytes": 245804,
+    "sha256": "f31e8572ef2509e4537328d1b080e92f9a4d833a859068ce991220df790cf884"
+  },
+  {
+    "path": "audio/orpheus/adapted/prompt_02.wav",
+    "bytes": 196652,
+    "sha256": "4ec5e55e4aa7b085f61961340a701dd2e784787e98ef4d03111a2fddc1aab457"
+  },
+  {
+    "path": "audio/orpheus/adapted/prompt_03.wav",
+    "bytes": 188460,
+    "sha256": "07e7c5b0b9d37cc2e87cc70c31dee0adf55db8e76772e28661a33d8b14f1afe9"
+  },
+  {
+    "path": "audio/orpheus/adapted/prompt_04.wav",
+    "bytes": 155692,
+    "sha256": "9033e23dcb4a31f4deb2108b04674b811c7c3c8a4b3fc638e8076675fec428ce"
+  },
+  {
+    "path": "audio/orpheus/adapted/prompt_05.wav",
+    "bytes": 180268,
+    "sha256": "1de6449d583fb48f0db9d1ba19bbfefda60016f32b246d9562d34a6a9bd7e322"
+  },
+  {
+    "path": "audio/orpheus/adapted/prompt_06.wav",
+    "bytes": 196652,
+    "sha256": "0b54348eb7b0d46cbcbd1869afa0b96fde6997d79036b19e11b2b7854261e468"
+  },
+  {
+    "path": "audio/orpheus/adapted/prompt_07.wav",
+    "bytes": 204844,
+    "sha256": "e71ca35cf0c12b1fb40d30cb26a285d0207ce9f845239c5ecfdd246e605f7903"
+  },
+  {
+    "path": "audio/orpheus/base/prompt_00.wav",
+    "bytes": 147500,
+    "sha256": "04e89b14b5aed1dfd9ccd11ba50eb8ab6d9d4282995ca48cebc77256ecb90216"
+  },
+  {
+    "path": "audio/orpheus/base/prompt_01.wav",
+    "bytes": 139308,
+    "sha256": "7980d7b59dccc9b2dea59581280823586b69ca1d48862d6fac2e74a7a9bfabc2"
+  },
+  {
+    "path": "audio/orpheus/base/prompt_02.wav",
+    "bytes": 221228,
+    "sha256": "b98b50173b0a0964eedcfd1286b9e258bf5c0d304b6dcd881ac10c8bf9709918"
+  },
+  {
+    "path": "audio/orpheus/base/prompt_03.wav",
+    "bytes": 180268,
+    "sha256": "b91552513a2ed17fab55c484e16cffd70d917a360b4735b2aea250845563b4c3"
+  },
+  {
+    "path": "audio/orpheus/base/prompt_04.wav",
+    "bytes": 163884,
+    "sha256": "45421df4a7cec956ec603fbb64811cc87fe6947afc3a0335dd8b32d40ec9ccd8"
+  },
+  {
+    "path": "audio/orpheus/base/prompt_05.wav",
+    "bytes": 208940,
+    "sha256": "a2431d6ef66a4a798a528d185d0707f430d608764acda9586e91f82215544afb"
+  },
+  {
+    "path": "audio/orpheus/base/prompt_06.wav",
+    "bytes": 184364,
+    "sha256": "16af5a42ac769e4194002ea1465f0666e0990d4a231e370808cc2f9057665b7d"
+  },
+  {
+    "path": "audio/orpheus/base/prompt_07.wav",
+    "bytes": 196652,
+    "sha256": "143f28842c83448b8fcea73017526274be7ac061f0aa3b727ea623f9839bcaad"
+  },
+  {
+    "path": "audio/sesame/adapted/giggle_0_neutral.wav",
+    "bytes": 96044,
+    "sha256": "16cbff670bebf59188345a79fcc99952ee5f7ba925e214874c71399b8b684fb6"
+  },
+  {
+    "path": "audio/sesame/adapted/giggle_0_tagged.wav",
+    "bytes": 130604,
+    "sha256": "e0d7ef91ff0c52b71a8630aa598bfbc30971149f36915fd3252511cbeb3d9a2c"
+  },
+  {
+    "path": "audio/sesame/adapted/giggle_1_neutral.wav",
+    "bytes": 157484,
+    "sha256": "653c1ed9cf6e11412cc935dee7dff90f0d059a72137faec5a63c795459658e15"
+  },
+  {
+    "path": "audio/sesame/adapted/giggle_1_tagged.wav",
+    "bytes": 234284,
+    "sha256": "9b72e5252d8b6f7a6f4c7ec3bbbf0f45e39bb2a9e6a16db8e140a9c6e9f32c2c"
+  },
+  {
+    "path": "audio/sesame/adapted/laugh_0_neutral.wav",
+    "bytes": 188204,
+    "sha256": "464bbcf2b6292d8a259cbb594e3c14a0e5ba8ca531bcac391e6a6495f9a7420c"
+  },
+  {
+    "path": "audio/sesame/adapted/laugh_0_tagged.wav",
+    "bytes": 172844,
+    "sha256": "00cf9a83ce1694de4ddab137b668306acce45c5459cee81fa3702a3a5cc94359"
+  },
+  {
+    "path": "audio/sesame/adapted/laugh_1_neutral.wav",
+    "bytes": 119084,
+    "sha256": "4cba912639254c066d906b318531574ed1c2d664c2ece8e8b2f185fa85298e88"
+  },
+  {
+    "path": "audio/sesame/adapted/laugh_1_tagged.wav",
+    "bytes": 180524,
+    "sha256": "0b25fc0881603af5831526dd4fcfcfb7dfe1539b3185ddd4358b230d16b50ea8"
+  },
+  {
+    "path": "audio/sesame/adapted/sigh_0_neutral.wav",
+    "bytes": 130604,
+    "sha256": "1cbb54147d1375513821f170b229379c8285d264d7ba4b4c568fcdb75bf3c7f0"
+  },
+  {
+    "path": "audio/sesame/adapted/sigh_0_tagged.wav",
+    "bytes": 188204,
+    "sha256": "d432582bfebcd4961ca2d3dfcfd944df252ed4826172f5987325172bcb23fd4e"
+  },
+  {
+    "path": "audio/sesame/adapted/sigh_1_neutral.wav",
+    "bytes": 111404,
+    "sha256": "b00babd0052826c46bc90a4d1e5335069ad26f6670ad7e90669f71320f715a8a"
+  },
+  {
+    "path": "audio/sesame/adapted/sigh_1_tagged.wav",
+    "bytes": 161324,
+    "sha256": "6da2951a4629a2a809e8595bd3940d5da3fe494e4e89cec5c728dcc0fd9d22d9"
+  },
+  {
+    "path": "audio/sesame/base/giggle_0_neutral.wav",
+    "bytes": 80684,
+    "sha256": "b3a6df7fe20a4e713f19f961772e908477962b335419ac5289b05fea2d9adebd"
+  },
+  {
+    "path": "audio/sesame/base/giggle_0_tagged.wav",
+    "bytes": 142124,
+    "sha256": "fe8a33bf80d471d1812c6d37586e5989eadcc10ae61c48ed02e4e516eafd2dc3"
+  },
+  {
+    "path": "audio/sesame/base/giggle_1_neutral.wav",
+    "bytes": 142124,
+    "sha256": "73b02cd34995eb75549fd5941a67c1b621b7d4f613d140f83693955d9529cd44"
+  },
+  {
+    "path": "audio/sesame/base/giggle_1_tagged.wav",
+    "bytes": 284204,
+    "sha256": "c1f7efac3892a907c7e08ecef5c013fa83db23879a1a5d1d055f81ec7efd4348"
+  },
+  {
+    "path": "audio/sesame/base/laugh_0_neutral.wav",
+    "bytes": 130604,
+    "sha256": "beb4be0e62fa63ca5c81facb5f82f7a565e027f51e0df489d8f361664b49460f"
+  },
+  {
+    "path": "audio/sesame/base/laugh_0_tagged.wav",
+    "bytes": 288044,
+    "sha256": "1a475df9255e256277a1a4fc66fee780c604977381059a04f09b33f46ebc00ce"
+  },
+  {
+    "path": "audio/sesame/base/laugh_1_neutral.wav",
+    "bytes": 99884,
+    "sha256": "85f7ed5500add23a6d5a4e51df5c0c01e00ecb2c740863b0155cb22d547bf701"
+  },
+  {
+    "path": "audio/sesame/base/laugh_1_tagged.wav",
+    "bytes": 184364,
+    "sha256": "46f2c99ca4fd65fff0c724900e9baa1f7bfab25fbc6246c669c9678c463def96"
+  },
+  {
+    "path": "audio/sesame/base/sigh_0_neutral.wav",
+    "bytes": 111404,
+    "sha256": "661c54d5ffba92518ef690cade4d3338a7d5c6f8e67882bf024d557e2b5eeeff"
+  },
+  {
+    "path": "audio/sesame/base/sigh_0_tagged.wav",
+    "bytes": 288044,
+    "sha256": "90cbd9c00b4f27ce71723bc022d3a5a0503ea6039ed7456a2b50785d86986f30"
+  },
+  {
+    "path": "audio/sesame/base/sigh_1_neutral.wav",
+    "bytes": 107564,
+    "sha256": "eb79e2600f5c4c0d571ae6e6ba5dc26c84c1547a6fb5225a9467833789c61b8a"
+  },
+  {
+    "path": "audio/sesame/base/sigh_1_tagged.wav",
+    "bytes": 161324,
+    "sha256": "3e5b8de26910376f0906253ae1effff657b927efb6382a7ff4aaeea3080d0c58"
+  },
+  {
+    "path": "compatibility_failures.json",
+    "bytes": 2089,
+    "sha256": "b63d17feb8f4fa6d77b0930c27c2b5a11593c5f76e1e7681acbd4fc118f89725"
+  },
+  {
+    "path": "failure_comparisons.json",
+    "bytes": 1838,
+    "sha256": "4ca79650a568e6164c37fe2a76577b398245f9d372cbf7d640373d3eecfdb224"
+  },
+  {
+    "path": "orpheus_manifest.json",
+    "bytes": 9215,
+    "sha256": "ea5df497cb092ba6520cbe41d82209a0e688ae7f9ea74e7764132ff05b324d6a"
+  },
+  {
+    "path": "sesame_manifest.json",
+    "bytes": 12199,
+    "sha256": "86499cd9d2bcc89823432bb54b59072044bf232ccc47f34be5079e9289ed1542"
+  },
+  {
+    "path": "adapters/orpheus/adapter_model.safetensors",
+    "bytes": 97307544,
+    "sha256": "fbd43cd5287a4b69398c29f88286e60146e9328cb48e3c22e15a2a305b25b29f",
+    "storage": "huggingface",
+    "repository": "https://huggingface.co/bojieli/exp8-6-orpheus-elise-lora",
+    "revision": "536092e9479fa1717e2b8f9cc1be52728b273e95"
+  },
+  {
+    "path": "adapters/orpheus/tokenizer.json",
+    "bytes": 22849547,
+    "sha256": "fc3fecb199b4170636dbfab986d25f628157268d37b861f9cadaca60b1353bce",
+    "storage": "huggingface",
+    "repository": "https://huggingface.co/bojieli/exp8-6-orpheus-elise-lora",
+    "revision": "536092e9479fa1717e2b8f9cc1be52728b273e95"
+  },
+  {
+    "path": "adapters/orpheus/tokenizer_config.json",
+    "bytes": 5403483,
+    "sha256": "e2d66c40996da379756a3209c42283e7a8203b1238b54920b070ea384fdc3a9f",
+    "storage": "huggingface",
+    "repository": "https://huggingface.co/bojieli/exp8-6-orpheus-elise-lora",
+    "revision": "536092e9479fa1717e2b8f9cc1be52728b273e95"
+  },
+  {
+    "path": "adapters/sesame/adapter_model.safetensors",
+    "bytes": 58125112,
+    "sha256": "23994b467b319f99bbc642f4dc2ba4d39718838cd7cbfe949311a42ebcb5480e",
+    "storage": "huggingface",
+    "repository": "https://huggingface.co/bojieli/exp8-6-sesame-elise-tags-lora",
+    "revision": "f2e042be0f38d6078976ef7e16cf49b91097f756"
+  },
+  {
+    "path": "adapters/sesame/tokenizer.json",
+    "bytes": 17210159,
+    "sha256": "92487a7224af2e53dc421a3f73056078acea66f0e86734f8d41babce66bad5cc",
+    "storage": "huggingface",
+    "repository": "https://huggingface.co/bojieli/exp8-6-sesame-elise-tags-lora",
+    "revision": "f2e042be0f38d6078976ef7e16cf49b91097f756"
+  },
+  {
+    "path": "adapters/sesame/tokenizer_config.json",
+    "bytes": 50577,
+    "sha256": "eb03e82f89eebfd9b9e2c1a2741dfc34c8982c3c14d132b0ec42fed737b832d3",
+    "storage": "huggingface",
+    "repository": "https://huggingface.co/bojieli/exp8-6-sesame-elise-tags-lora",
+    "revision": "f2e042be0f38d6078976ef7e16cf49b91097f756"
+  }
+]

chapter8/speech-sft-experiment/validation/exp8-6-20260804-v1/compatibility_failures.json

@@ -0,0 +1,30 @@
+{
+  "experiment": "8-6",
+  "status": "resolved_without_reducing_campaign",
+  "failures": [
+    {
+      "stage": "dataset_load",
+      "component": "datasets 3.6.0 / Hugging Face Hub",
+      "attempted_source": "MrDragonFox/Elise",
+      "exception_type": "FileNotFoundError",
+      "message": "The dataset repository exists but Hugging Face marks it disabled, so load_dataset could not resolve its data files.",
+      "resolution": "Pinned the public non-disabled maxbsoft/mrdragonfox-elise mirror at revision 2cc657c3f94a83df18fcd968b7531ca1a19c7f88. Both training manifests record the substitution."
+    },
+    {
+      "stage": "sesame_model_load",
+      "component": "Unsloth 2026.8.2",
+      "exception_type": "RuntimeError",
+      "message": "Unsloth: Could not find a valid pad token for unsloth/csm-1b - please inspect the tokenizer. A temporary '<｜PAD▁TOKEN｜>' was added.",
+      "root_cause": "The upstream tokenizer aliases pad to EOS while CSM config declares existing token 128004 (<|finetune_right_pad_id|>) as padding; the 2026.8 safety guard also interprets CSM's audio-codebook vocab_size during validation.",
+      "resolution": "Used standard Transformers + PEFT LoRA for Sesame, with a temporary tokenizer metadata snapshot pointing pad_token to the existing model-configured token 128004. No vocabulary item or model weight was added."
+    },
+    {
+      "stage": "sesame_pre_training_evaluation",
+      "component": "Transformers 4.57.6 CsmForConditionalGeneration",
+      "exception_type": "RuntimeError",
+      "message": "Index put requires the source and destination dtypes match, got BFloat16 for the destination and Float for the source.",
+      "root_cause": "CSM's audio codec returned float32 audio embeddings while bf16 autocast produced bf16 text embedding slots in _merge_input_ids_with_input_values.",
+      "resolution": "Ran the 1B Sesame model and its LoRA campaign in float32 (bf16 disabled). The final campaign retained the same 168 training examples, 24 held-out examples, 60 optimizer steps, and 24 base/adapted comparison WAVs."
+    }
+  ]
+}

chapter8/speech-sft-experiment/validation/exp8-6-20260804-v1/failure_comparisons.json

@@ -0,0 +1,68 @@
+[
+  {
+    "track": "orpheus",
+    "arm": "base",
+    "reason": "lowest_cross_sentence_timbre_proxy",
+    "prompt_a": 1,
+    "prompt_b": 7,
+    "cosine": 0.9738007187843323
+  },
+  {
+    "track": "orpheus",
+    "arm": "adapted",
+    "reason": "lowest_cross_sentence_timbre_proxy",
+    "prompt_a": 1,
+    "prompt_b": 5,
+    "cosine": 0.9688550233840942
+  },
+  {
+    "track": "sesame",
+    "arm": "base",
+    "reason": "tag_did_not_raise_matching_audioset_score",
+    "pair_id": "giggle_0",
+    "tag": "giggle",
+    "neutral_score": 0.0009356054943054914,
+    "tagged_score": 0.0006116584991104901,
+    "tagged_minus_neutral": -0.00032394699519500136
+  },
+  {
+    "track": "sesame",
+    "arm": "base",
+    "reason": "tag_did_not_raise_matching_audioset_score",
+    "pair_id": "giggle_1",
+    "tag": "giggle",
+    "neutral_score": 0.0008636609418317676,
+    "tagged_score": 0.0005855335621163249,
+    "tagged_minus_neutral": -0.00027812737971544266
+  },
+  {
+    "track": "sesame",
+    "arm": "base",
+    "reason": "tag_did_not_raise_matching_audioset_score",
+    "pair_id": "sigh_1",
+    "tag": "sigh",
+    "neutral_score": 0.0024482603184878826,
+    "tagged_score": 0.0006605299422517419,
+    "tagged_minus_neutral": -0.0017877303762361407
+  },
+  {
+    "track": "sesame",
+    "arm": "adapted",
+    "reason": "tag_did_not_raise_matching_audioset_score",
+    "pair_id": "giggle_1",
+    "tag": "giggle",
+    "neutral_score": 0.003269059816375375,
+    "tagged_score": 0.0007443267968483269,
+    "tagged_minus_neutral": -0.002524733019527048
+  },
+  {
+    "track": "sesame",
+    "arm": "adapted",
+    "reason": "tag_did_not_raise_matching_audioset_score",
+    "pair_id": "laugh_1",
+    "tag": "laugh",
+    "neutral_score": 0.00033708245609886944,
+    "tagged_score": 0.0003236289485357702,
+    "tagged_minus_neutral": -1.3453507563099265e-05
+  }
+]

chapter8/speech-sft-experiment/validation/exp8-6-20260804-v1/orpheus_manifest.json

@@ -0,0 +1,264 @@
+{
+  "experiment": "8-6",
+  "track": "orpheus_cross_sentence_voice_consistency",
+  "status": "trained_and_generated",
+  "seed": 7601,
+  "base_model": "unsloth/orpheus-3b-0.1-ft",
+  "base_model_revision": "eae2b6e5e429c81b95ac42a883ac64f126583d43",
+  "dataset": "maxbsoft/mrdragonfox-elise",
+  "dataset_revision": "2cc657c3f94a83df18fcd968b7531ca1a19c7f88",
+  "source_dataset_note": "Public non-disabled mirror of the disabled MrDragonFox/Elise dataset named by the upstream notebook.",
+  "train_examples_requested": 128,
+  "train_examples_encoded": 128,
+  "eval_examples_requested": 16,
+  "eval_examples_encoded": 16,
+  "train_failures": [],
+  "eval_failures": [],
+  "max_audio_seconds": 4.0,
+  "optimizer_steps": 60,
+  "effective_batch_size": 4,
+  "lora_rank": 16,
+  "pre_eval": {
+    "eval_loss": 5.237792491912842,
+    "eval_model_preparation_time": 0.0093,
+    "eval_runtime": 0.7249,
+    "eval_samples_per_second": 22.072,
+    "eval_steps_per_second": 22.072
+  },
+  "train_metrics": {
+    "train_runtime": 24.1374,
+    "train_samples_per_second": 9.943,
+    "train_steps_per_second": 2.486,
+    "total_flos": 1303634560167936.0,
+    "train_loss": 4.880102078119914,
+    "epoch": 1.875
+  },
+  "post_eval": {
+    "eval_loss": 4.86582088470459,
+    "eval_model_preparation_time": 0.0093,
+    "eval_runtime": 0.3247,
+    "eval_samples_per_second": 49.269,
+    "eval_steps_per_second": 49.269,
+    "epoch": 1.875
+  },
+  "gpu": "NVIDIA RTX PRO 6000 Blackwell Workstation Edition",
+  "peak_gpu_memory_bytes": 9277800448,
+  "wall_seconds": 202.15462517738342,
+  "adapter_local_files": [
+    {
+      "path": "adapters/orpheus/README.md",
+      "bytes": 5214,
+      "sha256": "d62b94e4ab068c9379c9d6a92fa73dc9468f85ce204615543fa20db6e453e4fb"
+    },
+    {
+      "path": "adapters/orpheus/adapter_config.json",
+      "bytes": 1238,
+      "sha256": "40bd5beba0cbeb72bbe0acf0371d0ab64f1217722d3590ee165f5f9d2f745ca9"
+    },
+    {
+      "path": "adapters/orpheus/adapter_model.safetensors",
+      "bytes": 97307544,
+      "sha256": "fbd43cd5287a4b69398c29f88286e60146e9328cb48e3c22e15a2a305b25b29f"
+    },
+    {
+      "path": "adapters/orpheus/chat_template.jinja",
+      "bytes": 3827,
+      "sha256": "5816fce10444e03c2e9ee1ef8a4a1ea61ae7e69e438613f3b17b69d0426223a4"
+    },
+    {
+      "path": "adapters/orpheus/special_tokens_map.json",
+      "bytes": 508,
+      "sha256": "f17390e64274b3117a7f99afa87eb1de93097b820842c68bd626aff53d73d7f5"
+    },
+    {
+      "path": "adapters/orpheus/tokenizer.json",
+      "bytes": 22849547,
+      "sha256": "fc3fecb199b4170636dbfab986d25f628157268d37b861f9cadaca60b1353bce"
+    },
+    {
+      "path": "adapters/orpheus/tokenizer_config.json",
+      "bytes": 5403483,
+      "sha256": "e2d66c40996da379756a3209c42283e7a8203b1238b54920b070ea384fdc3a9f"
+    }
+  ],
+  "adapter_huggingface_repo": "https://huggingface.co/bojieli/exp8-6-orpheus-elise-lora",
+  "adapter_huggingface_revision": "536092e9479fa1717e2b8f9cc1be52728b273e95",
+  "audio": [
+    {
+      "prompt_id": 0,
+      "prompt": "The morning train crossed the bridge just before sunrise.",
+      "seed": 7601,
+      "path": "audio/orpheus/base/prompt_00.wav",
+      "sha256": "04e89b14b5aed1dfd9ccd11ba50eb8ab6d9d4282995ca48cebc77256ecb90216",
+      "samples": 73728,
+      "seconds": 3.072,
+      "decoded_frames": 36,
+      "first_invalid_frame": null
+    },
+    {
+      "prompt_id": 1,
+      "prompt": "Please leave the blue notebook beside the kitchen window.",
+      "seed": 7602,
+      "path": "audio/orpheus/base/prompt_01.wav",
+      "sha256": "7980d7b59dccc9b2dea59581280823586b69ca1d48862d6fac2e74a7a9bfabc2",
+      "samples": 69632,
+      "seconds": 2.9013333333333335,
+      "decoded_frames": 34,
+      "first_invalid_frame": null
+    },
+    {
+      "prompt_id": 2,
+      "prompt": "A patient astronomer mapped every bright star in the winter sky.",
+      "seed": 7603,
+      "path": "audio/orpheus/base/prompt_02.wav",
+      "sha256": "b98b50173b0a0964eedcfd1286b9e258bf5c0d304b6dcd881ac10c8bf9709918",
+      "samples": 110592,
+      "seconds": 4.608,
+      "decoded_frames": 54,
+      "first_invalid_frame": null
+    },
+    {
+      "prompt_id": 3,
+      "prompt": "We walked home slowly while the last shops turned off their lights.",
+      "seed": 7604,
+      "path": "audio/orpheus/base/prompt_03.wav",
+      "sha256": "b91552513a2ed17fab55c484e16cffd70d917a360b4735b2aea250845563b4c3",
+      "samples": 90112,
+      "seconds": 3.7546666666666666,
+      "decoded_frames": 44,
+      "first_invalid_frame": null
+    },
+    {
+      "prompt_id": 4,
+      "prompt": "Could you read the final paragraph one more time for the group?",
+      "seed": 7605,
+      "path": "audio/orpheus/base/prompt_04.wav",
+      "sha256": "45421df4a7cec956ec603fbb64811cc87fe6947afc3a0335dd8b32d40ec9ccd8",
+      "samples": 81920,
+      "seconds": 3.4133333333333336,
+      "decoded_frames": 40,
+      "first_invalid_frame": null
+    },
+    {
+      "prompt_id": 5,
+      "prompt": "The small garden stayed green even through the hottest week of July.",
+      "seed": 7606,
+      "path": "audio/orpheus/base/prompt_05.wav",
+      "sha256": "a2431d6ef66a4a798a528d185d0707f430d608764acda9586e91f82215544afb",
+      "samples": 104448,
+      "seconds": 4.352,
+      "decoded_frames": 51,
+      "first_invalid_frame": null
+    },
+    {
+      "prompt_id": 6,
+      "prompt": "I packed a warm coat, two apples, and a compass for the long hike.",
+      "seed": 7607,
+      "path": "audio/orpheus/base/prompt_06.wav",
+      "sha256": "16af5a42ac769e4194002ea1465f0666e0990d4a231e370808cc2f9057665b7d",
+      "samples": 92160,
+      "seconds": 3.84,
+      "decoded_frames": 45,
+      "first_invalid_frame": null
+    },
+    {
+      "prompt_id": 7,
+      "prompt": "Tomorrow's meeting begins at nine, so I will arrive a little early.",
+      "seed": 7608,
+      "path": "audio/orpheus/base/prompt_07.wav",
+      "sha256": "143f28842c83448b8fcea73017526274be7ac061f0aa3b727ea623f9839bcaad",
+      "samples": 98304,
+      "seconds": 4.096,
+      "decoded_frames": 48,
+      "first_invalid_frame": null
+    },
+    {
+      "prompt_id": 0,
+      "prompt": "The morning train crossed the bridge just before sunrise.",
+      "seed": 7601,
+      "path": "audio/orpheus/adapted/prompt_00.wav",
+      "sha256": "24470415a507117b9f2b6a4b3f13350a453febca782f542e592ea6d41678f9da",
+      "samples": 110592,
+      "seconds": 4.608,
+      "decoded_frames": 54,
+      "first_invalid_frame": null
+    },
+    {
+      "prompt_id": 1,
+      "prompt": "Please leave the blue notebook beside the kitchen window.",
+      "seed": 7602,
+      "path": "audio/orpheus/adapted/prompt_01.wav",
+      "sha256": "f31e8572ef2509e4537328d1b080e92f9a4d833a859068ce991220df790cf884",
+      "samples": 122880,
+      "seconds": 5.12,
+      "decoded_frames": 60,
+      "first_invalid_frame": null
+    },
+    {
+      "prompt_id": 2,
+      "prompt": "A patient astronomer mapped every bright star in the winter sky.",
+      "seed": 7603,
+      "path": "audio/orpheus/adapted/prompt_02.wav",
+      "sha256": "4ec5e55e4aa7b085f61961340a701dd2e784787e98ef4d03111a2fddc1aab457",
+      "samples": 98304,
+      "seconds": 4.096,
+      "decoded_frames": 48,
+      "first_invalid_frame": null
+    },
+    {
+      "prompt_id": 3,
+      "prompt": "We walked home slowly while the last shops turned off their lights.",
+      "seed": 7604,
+      "path": "audio/orpheus/adapted/prompt_03.wav",
+      "sha256": "07e7c5b0b9d37cc2e87cc70c31dee0adf55db8e76772e28661a33d8b14f1afe9",
+      "samples": 94208,
+      "seconds": 3.925333333333333,
+      "decoded_frames": 46,
+      "first_invalid_frame": null
+    },
+    {
+      "prompt_id": 4,
+      "prompt": "Could you read the final paragraph one more time for the group?",
+      "seed": 7605,
+      "path": "audio/orpheus/adapted/prompt_04.wav",
+      "sha256": "9033e23dcb4a31f4deb2108b04674b811c7c3c8a4b3fc638e8076675fec428ce",
+      "samples": 77824,
+      "seconds": 3.2426666666666666,
+      "decoded_frames": 38,
+      "first_invalid_frame": null
+    },
+    {
+      "prompt_id": 5,
+      "prompt": "The small garden stayed green even through the hottest week of July.",
+      "seed": 7606,
+      "path": "audio/orpheus/adapted/prompt_05.wav",
+      "sha256": "1de6449d583fb48f0db9d1ba19bbfefda60016f32b246d9562d34a6a9bd7e322",
+      "samples": 90112,
+      "seconds": 3.7546666666666666,
+      "decoded_frames": 44,
+      "first_invalid_frame": null
+    },
+    {
+      "prompt_id": 6,
+      "prompt": "I packed a warm coat, two apples, and a compass for the long hike.",
+      "seed": 7607,
+      "path": "audio/orpheus/adapted/prompt_06.wav",
+      "sha256": "0b54348eb7b0d46cbcbd1869afa0b96fde6997d79036b19e11b2b7854261e468",
+      "samples": 98304,
+      "seconds": 4.096,
+      "decoded_frames": 48,
+      "first_invalid_frame": null
+    },
+    {
+      "prompt_id": 7,
+      "prompt": "Tomorrow's meeting begins at nine, so I will arrive a little early.",
+      "seed": 7608,
+      "path": "audio/orpheus/adapted/prompt_07.wav",
+      "sha256": "e71ca35cf0c12b1fb40d30cb26a285d0207ce9f845239c5ecfdd246e605f7903",
+      "samples": 102400,
+      "seconds": 4.266666666666667,
+      "decoded_frames": 50,
+      "first_invalid_frame": null
+    }
+  ]
+}

chapter8/speech-sft-experiment/validation/exp8-6-20260804-v1/sesame_manifest.json

@@ -0,0 +1,369 @@
+{
+  "experiment": "8-6",
+  "track": "sesame_paralinguistic_tags",
+  "status": "trained_and_generated",
+  "seed": 7602,
+  "base_model": "unsloth/csm-1b",
+  "base_model_revision": "39e8c756bec133ad7eb9ea1097c84a2bd891c949",
+  "dataset": "maxbsoft/mrdragonfox-elise",
+  "dataset_revision": "2cc657c3f94a83df18fcd968b7531ca1a19c7f88",
+  "source_dataset_note": "Public non-disabled mirror of the disabled MrDragonFox/Elise dataset named by the upstream notebook.",
+  "train_examples_selected": 168,
+  "train_examples_preprocessed": 168,
+  "train_category_counts": {
+    "laugh": 48,
+    "giggle": 32,
+    "sigh": 48,
+    "neutral": 40
+  },
+  "eval_examples_selected": 24,
+  "eval_examples_preprocessed": 24,
+  "eval_category_counts": {
+    "laugh": 8,
+    "giggle": 0,
+    "sigh": 8,
+    "neutral": 8
+  },
+  "train_failures": [],
+  "eval_failures": [],
+  "max_audio_seconds": 8.0,
+  "optimizer_steps": 60,
+  "effective_batch_size": 4,
+  "lora_rank": 16,
+  "pre_eval": {
+    "eval_loss": 128.2307586669922,
+    "eval_model_preparation_time": 0.0114,
+    "eval_runtime": 1.3794,
+    "eval_samples_per_second": 17.398,
+    "eval_steps_per_second": 17.398
+  },
+  "train_metrics": {
+    "train_runtime": 26.4676,
+    "train_samples_per_second": 9.068,
+    "train_steps_per_second": 2.267,
+    "total_flos": 460628125655040.0,
+    "train_loss": 124.45613021850586,
+    "epoch": 1.4285714285714286
+  },
+  "post_eval": {
+    "eval_loss": 124.34239959716797,
+    "eval_model_preparation_time": 0.0114,
+    "eval_runtime": 1.8573,
+    "eval_samples_per_second": 12.922,
+    "eval_steps_per_second": 12.922,
+    "epoch": 1.4285714285714286
+  },
+  "gpu": "NVIDIA RTX PRO 6000 Blackwell Workstation Edition",
+  "peak_gpu_memory_bytes": 10242490368,
+  "wall_seconds": 182.60698008537292,
+  "adapter_local_files": [
+    {
+      "path": "adapters/sesame/README.md",
+      "bytes": 5183,
+      "sha256": "65a8c53feecbfc1e7a8cd64bc6a79baedbffd1ef6b9cc17b1ea0a5529e4d74ec"
+    },
+    {
+      "path": "adapters/sesame/adapter_config.json",
+      "bytes": 1091,
+      "sha256": "44c3053b5037214ea8862003dfeea867554c9fbb83df5728ea80e9a0915e3ee2"
+    },
+    {
+      "path": "adapters/sesame/adapter_model.safetensors",
+      "bytes": 58125112,
+      "sha256": "23994b467b319f99bbc642f4dc2ba4d39718838cd7cbfe949311a42ebcb5480e"
+    },
+    {
+      "path": "adapters/sesame/chat_template.jinja",
+      "bytes": 2002,
+      "sha256": "b43753fbbc23be93160cccf825d2cf97dec39538cb274350216940e7acbc0fe7"
+    },
+    {
+      "path": "adapters/sesame/preprocessor_config.json",
+      "bytes": 271,
+      "sha256": "8648040cda0469976874c733acbec88c0ef094cd95628ab63b32950f01872af9"
+    },
+    {
+      "path": "adapters/sesame/special_tokens_map.json",
+      "bytes": 459,
+      "sha256": "32f404d626cf7b1b6eea36c241ae7cbd6ec29c777a8701ea48c0fe9ebb94c9b1"
+    },
+    {
+      "path": "adapters/sesame/tokenizer.json",
+      "bytes": 17210159,
+      "sha256": "92487a7224af2e53dc421a3f73056078acea66f0e86734f8d41babce66bad5cc"
+    },
+    {
+      "path": "adapters/sesame/tokenizer_config.json",
+      "bytes": 50577,
+      "sha256": "eb03e82f89eebfd9b9e2c1a2741dfc34c8982c3c14d132b0ec42fed737b832d3"
+    }
+  ],
+  "adapter_huggingface_repo": "https://huggingface.co/bojieli/exp8-6-sesame-elise-tags-lora",
+  "adapter_huggingface_revision": "f2e042be0f38d6078976ef7e16cf49b91097f756",
+  "audio": [
+    {
+      "pair_id": "laugh_0",
+      "tag": "laugh",
+      "condition": "neutral",
+      "text": "I finally found the missing keys in my other pocket.",
+      "seed": 7602,
+      "path": "audio/sesame/base/laugh_0_neutral.wav",
+      "sha256": "beb4be0e62fa63ca5c81facb5f82f7a565e027f51e0df489d8f361664b49460f",
+      "samples": 65280,
+      "seconds": 2.72
+    },
+    {
+      "pair_id": "laugh_0",
+      "tag": "laugh",
+      "condition": "tagged",
+      "text": "I finally found the missing keys <laughs> in my other pocket.",
+      "seed": 7602,
+      "path": "audio/sesame/base/laugh_0_tagged.wav",
+      "sha256": "1a475df9255e256277a1a4fc66fee780c604977381059a04f09b33f46ebc00ce",
+      "samples": 144000,
+      "seconds": 6.0
+    },
+    {
+      "pair_id": "laugh_1",
+      "tag": "laugh",
+      "condition": "neutral",
+      "text": "That was the strangest joke I heard all week.",
+      "seed": 7603,
+      "path": "audio/sesame/base/laugh_1_neutral.wav",
+      "sha256": "85f7ed5500add23a6d5a4e51df5c0c01e00ecb2c740863b0155cb22d547bf701",
+      "samples": 49920,
+      "seconds": 2.08
+    },
+    {
+      "pair_id": "laugh_1",
+      "tag": "laugh",
+      "condition": "tagged",
+      "text": "That was the strangest joke <laughs> I heard all week.",
+      "seed": 7603,
+      "path": "audio/sesame/base/laugh_1_tagged.wav",
+      "sha256": "46f2c99ca4fd65fff0c724900e9baa1f7bfab25fbc6246c669c9678c463def96",
+      "samples": 92160,
+      "seconds": 3.84
+    },
+    {
+      "pair_id": "giggle_0",
+      "tag": "giggle",
+      "condition": "neutral",
+      "text": "You remembered the secret code after all.",
+      "seed": 7604,
+      "path": "audio/sesame/base/giggle_0_neutral.wav",
+      "sha256": "b3a6df7fe20a4e713f19f961772e908477962b335419ac5289b05fea2d9adebd",
+      "samples": 40320,
+      "seconds": 1.68
+    },
+    {
+      "pair_id": "giggle_0",
+      "tag": "giggle",
+      "condition": "tagged",
+      "text": "You remembered the secret code <giggles> after all.",
+      "seed": 7604,
+      "path": "audio/sesame/base/giggle_0_tagged.wav",
+      "sha256": "fe8a33bf80d471d1812c6d37586e5989eadcc10ae61c48ed02e4e516eafd2dc3",
+      "samples": 71040,
+      "seconds": 2.96
+    },
+    {
+      "pair_id": "giggle_1",
+      "tag": "giggle",
+      "condition": "neutral",
+      "text": "The tiny puppy tried to carry the enormous slipper.",
+      "seed": 7605,
+      "path": "audio/sesame/base/giggle_1_neutral.wav",
+      "sha256": "73b02cd34995eb75549fd5941a67c1b621b7d4f613d140f83693955d9529cd44",
+      "samples": 71040,
+      "seconds": 2.96
+    },
+    {
+      "pair_id": "giggle_1",
+      "tag": "giggle",
+      "condition": "tagged",
+      "text": "The tiny puppy <giggles> tried to carry the enormous slipper.",
+      "seed": 7605,
+      "path": "audio/sesame/base/giggle_1_tagged.wav",
+      "sha256": "c1f7efac3892a907c7e08ecef5c013fa83db23879a1a5d1d055f81ec7efd4348",
+      "samples": 142080,
+      "seconds": 5.92
+    },
+    {
+      "pair_id": "sigh_0",
+      "tag": "sigh",
+      "condition": "neutral",
+      "text": "The last bus left before we reached the corner.",
+      "seed": 7606,
+      "path": "audio/sesame/base/sigh_0_neutral.wav",
+      "sha256": "661c54d5ffba92518ef690cade4d3338a7d5c6f8e67882bf024d557e2b5eeeff",
+      "samples": 55680,
+      "seconds": 2.32
+    },
+    {
+      "pair_id": "sigh_0",
+      "tag": "sigh",
+      "condition": "tagged",
+      "text": "The last bus left <sighs> before we reached the corner.",
+      "seed": 7606,
+      "path": "audio/sesame/base/sigh_0_tagged.wav",
+      "sha256": "90cbd9c00b4f27ce71723bc022d3a5a0503ea6039ed7456a2b50785d86986f30",
+      "samples": 144000,
+      "seconds": 6.0
+    },
+    {
+      "pair_id": "sigh_1",
+      "tag": "sigh",
+      "condition": "neutral",
+      "text": "I suppose we need to finish the paperwork again.",
+      "seed": 7607,
+      "path": "audio/sesame/base/sigh_1_neutral.wav",
+      "sha256": "eb79e2600f5c4c0d571ae6e6ba5dc26c84c1547a6fb5225a9467833789c61b8a",
+      "samples": 53760,
+      "seconds": 2.24
+    },
+    {
+      "pair_id": "sigh_1",
+      "tag": "sigh",
+      "condition": "tagged",
+      "text": "I suppose <sighs> we need to finish the paperwork again.",
+      "seed": 7607,
+      "path": "audio/sesame/base/sigh_1_tagged.wav",
+      "sha256": "3e5b8de26910376f0906253ae1effff657b927efb6382a7ff4aaeea3080d0c58",
+      "samples": 80640,
+      "seconds": 3.36
+    },
+    {
+      "pair_id": "laugh_0",
+      "tag": "laugh",
+      "condition": "neutral",
+      "text": "I finally found the missing keys in my other pocket.",
+      "seed": 7602,
+      "path": "audio/sesame/adapted/laugh_0_neutral.wav",
+      "sha256": "464bbcf2b6292d8a259cbb594e3c14a0e5ba8ca531bcac391e6a6495f9a7420c",
+      "samples": 94080,
+      "seconds": 3.92
+    },
+    {
+      "pair_id": "laugh_0",
+      "tag": "laugh",
+      "condition": "tagged",
+      "text": "I finally found the missing keys <laughs> in my other pocket.",
+      "seed": 7602,
+      "path": "audio/sesame/adapted/laugh_0_tagged.wav",
+      "sha256": "00cf9a83ce1694de4ddab137b668306acce45c5459cee81fa3702a3a5cc94359",
+      "samples": 86400,
+      "seconds": 3.6
+    },
+    {
+      "pair_id": "laugh_1",
+      "tag": "laugh",
+      "condition": "neutral",
+      "text": "That was the strangest joke I heard all week.",
+      "seed": 7603,
+      "path": "audio/sesame/adapted/laugh_1_neutral.wav",
+      "sha256": "4cba912639254c066d906b318531574ed1c2d664c2ece8e8b2f185fa85298e88",
+      "samples": 59520,
+      "seconds": 2.48
+    },
+    {
+      "pair_id": "laugh_1",
+      "tag": "laugh",
+      "condition": "tagged",
+      "text": "That was the strangest joke <laughs> I heard all week.",
+      "seed": 7603,
+      "path": "audio/sesame/adapted/laugh_1_tagged.wav",
+      "sha256": "0b25fc0881603af5831526dd4fcfcfb7dfe1539b3185ddd4358b230d16b50ea8",
+      "samples": 90240,
+      "seconds": 3.76
+    },
+    {
+      "pair_id": "giggle_0",
+      "tag": "giggle",
+      "condition": "neutral",
+      "text": "You remembered the secret code after all.",
+      "seed": 7604,
+      "path": "audio/sesame/adapted/giggle_0_neutral.wav",
+      "sha256": "16cbff670bebf59188345a79fcc99952ee5f7ba925e214874c71399b8b684fb6",
+      "samples": 48000,
+      "seconds": 2.0
+    },
+    {
+      "pair_id": "giggle_0",
+      "tag": "giggle",
+      "condition": "tagged",
+      "text": "You remembered the secret code <giggles> after all.",
+      "seed": 7604,
+      "path": "audio/sesame/adapted/giggle_0_tagged.wav",
+      "sha256": "e0d7ef91ff0c52b71a8630aa598bfbc30971149f36915fd3252511cbeb3d9a2c",
+      "samples": 65280,
+      "seconds": 2.72
+    },
+    {
+      "pair_id": "giggle_1",
+      "tag": "giggle",
+      "condition": "neutral",
+      "text": "The tiny puppy tried to carry the enormous slipper.",
+      "seed": 7605,
+      "path": "audio/sesame/adapted/giggle_1_neutral.wav",
+      "sha256": "653c1ed9cf6e11412cc935dee7dff90f0d059a72137faec5a63c795459658e15",
+      "samples": 78720,
+      "seconds": 3.28
+    },
+    {
+      "pair_id": "giggle_1",
+      "tag": "giggle",
+      "condition": "tagged",
+      "text": "The tiny puppy <giggles> tried to carry the enormous slipper.",
+      "seed": 7605,
+      "path": "audio/sesame/adapted/giggle_1_tagged.wav",
+      "sha256": "9b72e5252d8b6f7a6f4c7ec3bbbf0f45e39bb2a9e6a16db8e140a9c6e9f32c2c",
+      "samples": 117120,
+      "seconds": 4.88
+    },
+    {
+      "pair_id": "sigh_0",
+      "tag": "sigh",
+      "condition": "neutral",
+      "text": "The last bus left before we reached the corner.",
+      "seed": 7606,
+      "path": "audio/sesame/adapted/sigh_0_neutral.wav",
+      "sha256": "1cbb54147d1375513821f170b229379c8285d264d7ba4b4c568fcdb75bf3c7f0",
+      "samples": 65280,
+      "seconds": 2.72
+    },
+    {
+      "pair_id": "sigh_0",
+      "tag": "sigh",
+      "condition": "tagged",
+      "text": "The last bus left <sighs> before we reached the corner.",
+      "seed": 7606,
+      "path": "audio/sesame/adapted/sigh_0_tagged.wav",
+      "sha256": "d432582bfebcd4961ca2d3dfcfd944df252ed4826172f5987325172bcb23fd4e",
+      "samples": 94080,
+      "seconds": 3.92
+    },
+    {
+      "pair_id": "sigh_1",
+      "tag": "sigh",
+      "condition": "neutral",
+      "text": "I suppose we need to finish the paperwork again.",
+      "seed": 7607,
+      "path": "audio/sesame/adapted/sigh_1_neutral.wav",
+      "sha256": "b00babd0052826c46bc90a4d1e5335069ad26f6670ad7e90669f71320f715a8a",
+      "samples": 55680,
+      "seconds": 2.32
+    },
+    {
+      "pair_id": "sigh_1",
+      "tag": "sigh",
+      "condition": "tagged",
+      "text": "I suppose <sighs> we need to finish the paperwork again.",
+      "seed": 7607,
+      "path": "audio/sesame/adapted/sigh_1_tagged.wav",
+      "sha256": "6da2951a4629a2a809e8595bd3940d5da3fe494e4e89cec5c728dcc0fd9d22d9",
+      "samples": 80640,
+      "seconds": 3.36
+    }
+  ]
+}