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

chapter8/exact-copy-sft/validation/manifest.json

@@ -0,0 +1,101 @@
+{
+  "experiment": "exact-copy-sft",
+  "model": "Qwen/Qwen3-8B",
+  "model_open_weight": true,
+  "hardware": "NVIDIA RTX PRO 6000 Blackwell Workstation Edition",
+  "data_sha256": {
+    "boundary.jsonl": "25a1dab24c0ead98ba773648abb99143abdba21d2403127210656ba65417ccf7",
+    "eval.jsonl": "3ad8d96b12fc9c5bb1f0e3b040619be2b87b6d686027e13eed8783fb85598709",
+    "train.jsonl": "a56d5acc6cd24b8f6180f1043e01c84ff9d5d1351dfb3a323615fe442c8103ad"
+  },
+  "training_receipt": "validation/train_20260808T035121Z/training_receipt.json",
+  "evaluations": {
+    "eval_adapted_boundary.json": {
+      "label": "adapted",
+      "split": "boundary",
+      "count": 256,
+      "byte_exact": 0.80078125,
+      "mean_first_diff_byte": 54.19921875
+    },
+    "eval_adapted_eval.json": {
+      "label": "adapted",
+      "split": "eval",
+      "count": 256,
+      "byte_exact": 0.7890625,
+      "mean_first_diff_byte": 54.01953125
+    },
+    "eval_base_eval.json": {
+      "label": "base",
+      "split": "eval",
+      "count": 256,
+      "byte_exact": 0.375,
+      "mean_first_diff_byte": 29.015625
+    }
+  },
+  "tokenizer_audit": {
+    "probe_count": 512,
+    "tokenizers": {
+      "Qwen/Qwen3-8B": {
+        "vocab_size": 151669,
+        "roundtrip_rate": 0.80078125,
+        "mean_tokens": 34.603515625,
+        "failures": [
+          "eval-004",
+          "eval-009",
+          "eval-014",
+          "eval-019",
+          "eval-024",
+          "eval-029",
+          "eval-034",
+          "eval-039",
+          "eval-044",
+          "eval-049",
+          "eval-054",
+          "eval-059",
+          "eval-064",
+          "eval-069",
+          "eval-074",
+          "eval-079",
+          "eval-084",
+          "eval-089",
+          "eval-094",
+          "eval-099"
+        ]
+      },
+      "Qwen/Qwen2.5-0.5B-Instruct": {
+        "vocab_size": 151665,
+        "roundtrip_rate": 0.80078125,
+        "mean_tokens": 34.603515625,
+        "failures": [
+          "eval-004",
+          "eval-009",
+          "eval-014",
+          "eval-019",
+          "eval-024",
+          "eval-029",
+          "eval-034",
+          "eval-039",
+          "eval-044",
+          "eval-049",
+          "eval-054",
+          "eval-059",
+          "eval-064",
+          "eval-069",
+          "eval-074",
+          "eval-079",
+          "eval-084",
+          "eval-089",
+          "eval-094",
+          "eval-099"
+        ]
+      },
+      "mistralai/Mistral-7B-v0.1": {
+        "vocab_size": 32000,
+        "roundtrip_rate": 1.0,
+        "mean_tokens": 39.26171875,
+        "failures": []
+      }
+    }
+  },
+  "manual_audit": "validation/manual_audit.md"
+}

chapter8/exact-copy-sft/validation/manual_audit.md

@@ -0,0 +1,10 @@
+# 特殊字符串合成数据人工抽查记录
+
+本轮扩容后按 `tool_json`、`decoy_copy`、`verbatim` 三类任务、10 种语言上下文和 8 种文章体裁分层抽查 train/eval/boundary 的样本（每类每 split 至少 4 条，共 36 条）。逐条检查：
+
+1. `source` 与直接复述任务的 `target` 是否完全一致；
+2. `tool_json` 的 JSON 是否可解析，且 `old_string` 是否与 `source` 完全相同；
+3. 空格、换行、字面量 `\\n`/`\\t`、Unicode `é`、组合字符、中文和零宽字符是否按原样保存；
+4. decoy 是否与 TARGET 等长但内容不同，避免模型靠长度捷径作答。
+
+抽查未发现目标字符串漂移、JSON 结构错误或 decoy 标记歧义。全量确定性断言见 `tests/`，跨 tokenizer 的 encode→decode 结果见 [`tokenizer_audit.json`](tokenizer_audit.json)。

chapter8/exact-copy-sft/validation/tokenizer_audit.json

@@ -0,0 +1,65 @@
+{
+  "probe_count": 512,
+  "tokenizers": {
+    "Qwen/Qwen3-8B": {
+      "vocab_size": 151669,
+      "roundtrip_rate": 0.80078125,
+      "mean_tokens": 34.603515625,
+      "failures": [
+        "eval-004",
+        "eval-009",
+        "eval-014",
+        "eval-019",
+        "eval-024",
+        "eval-029",
+        "eval-034",
+        "eval-039",
+        "eval-044",
+        "eval-049",
+        "eval-054",
+        "eval-059",
+        "eval-064",
+        "eval-069",
+        "eval-074",
+        "eval-079",
+        "eval-084",
+        "eval-089",
+        "eval-094",
+        "eval-099"
+      ]
+    },
+    "Qwen/Qwen2.5-0.5B-Instruct": {
+      "vocab_size": 151665,
+      "roundtrip_rate": 0.80078125,
+      "mean_tokens": 34.603515625,
+      "failures": [
+        "eval-004",
+        "eval-009",
+        "eval-014",
+        "eval-019",
+        "eval-024",
+        "eval-029",
+        "eval-034",
+        "eval-039",
+        "eval-044",
+        "eval-049",
+        "eval-054",
+        "eval-059",
+        "eval-064",
+        "eval-069",
+        "eval-074",
+        "eval-079",
+        "eval-084",
+        "eval-089",
+        "eval-094",
+        "eval-099"
+      ]
+    },
+    "mistralai/Mistral-7B-v0.1": {
+      "vocab_size": 32000,
+      "roundtrip_rate": 1.0,
+      "mean_tokens": 39.26171875,
+      "failures": []
+    }
+  }
+}

chapter8/exact-copy-sft/validation/train_20260808T035121Z/training_receipt.json

@@ -0,0 +1,21 @@
+{
+  "experiment": "8-19-exact-copy-sft",
+  "run": "train_20260808T035121Z",
+  "model": "Qwen/Qwen3-8B",
+  "data_sha256": "a56d5acc6cd24b8f6180f1043e01c84ff9d5d1351dfb3a323615fe442c8103ad",
+  "train_examples": 1024,
+  "config": {
+    "model": "Qwen/Qwen3-8B",
+    "epochs": 2,
+    "lr": 0.0001,
+    "batch_size": 2,
+    "grad_accum": 4,
+    "max_length": 768,
+    "seed": 719,
+    "output": "/home/ubuntu/ai-agent-book/chapter8/exact-copy-sft/output/adapter"
+  },
+  "cuda": "NVIDIA RTX PRO 6000 Blackwell Workstation Edition",
+  "started_at": "2026-08-08T03:51:21.085606+00:00",
+  "finished_at": "2026-08-08T03:53:11.442422+00:00",
+  "final_loss": 1.6291709243887453e-06
+}