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/generate_data.py

@@ -0,0 +1,60 @@
+from __future__ import annotations
+import hashlib, json, random
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parent
+DATA = ROOT / "data"
+ALPHABET = "ABCDEFGHJKLMNPQRSTUVWXYZabcdefghijkmnopqrstuvwxyz23456789_-+=:@"
+LANGUAGES = ["Python", "JavaScript", "Java", "Go", "Rust", "SQL", "Shell", "JSON", "Markdown", "纯文本"]
+ARTICLE_TYPES = ["代码审查", "补丁说明", "故障复盘", "API 文档", "部署手册", "测试报告", "变更公告", "设计文档"]
+
+def make_string(rng, n):
+    return "".join(rng.choice(ALPHABET) for _ in range(n))
+
+def make(split, n, seed):
+    rng = random.Random(seed)
+    rows = []
+    for i in range(n):
+        s = make_string(rng, 18 + (i % 7) * 7)
+        if i % 5 == 0:
+            s = s[:8] + " " + s[8:] + "\n" + s[-3:]
+        elif i % 5 == 1:
+            s = s[:5] + "\\n" + s[5:] + "\\tEND"
+        elif i % 5 == 2:
+            s = "ID=" + s + "; sha256=" + make_string(rng, 16)
+        elif i % 5 == 3:
+            s = s[:4] + "é" + s[4:] + "\u200b"
+        else:
+            s = "中" + s + "e\u0301"
+        candidates = [make_string(rng, len(s)), make_string(rng, len(s)), s]
+        language = LANGUAGES[(i + seed) % len(LANGUAGES)]
+        article_type = ARTICLE_TYPES[(i * 3 + seed) % len(ARTICLE_TYPES)]
+        if i % 3 == 0:
+            prompt = (f"这是{article_type}中的 {language} 文件。\n从下面的工具观察中找到 path=/tmp/job.txt 的 old_string。\n"
+                      "只输出 JSON，不要解释：\n" + json.dumps(
+                          {"path": "/tmp/job.txt", "old_string": s, "other": ""}, ensure_ascii=False))
+            target = json.dumps({"old_string": s}, ensure_ascii=False, separators=(",", ":"))
+            kind = "tool_json"
+        elif i % 3 == 1:
+            prompt = (f"这是{article_type}中的 {language} 片段。\n请逐字复述唯一被标记为 TARGET 的字符串，只输出字符串本身。\n" +
+                      "\n".join(f"{('TARGET' if j == 2 else 'DECOY')}={x}" for j, x in enumerate(candidates)))
+            target, kind = s, "decoy_copy"
+        else:
+            prompt = (f"这是{article_type}中的 {language} 片段。\n请把 SOURCE 中的内容逐字复制到 ANSWER，不得修正大小写、空格、反斜杠或换行，只输出 ANSWER。\nSOURCE:\n" + s)
+            target, kind = s, "verbatim"
+        rows.append({"id": f"{split}-{i:03d}", "kind": kind, "language": language,
+                     "article_type": article_type, "source": s, "prompt": prompt, "target": target})
+    return rows
+
+def write(name, rows):
+    p = DATA / name
+    p.parent.mkdir(parents=True, exist_ok=True)
+    p.write_text("\n".join(json.dumps(x, ensure_ascii=False) for x in rows) + "\n", encoding="utf-8")
+    return hashlib.sha256(p.read_bytes()).hexdigest()
+
+if __name__ == "__main__":
+    print(json.dumps({
+        "train": write("train.jsonl", make("train", 1024, 719)),
+        "eval": write("eval.jsonl", make("eval", 256, 1729)),
+        "boundary": write("boundary.jsonl", make("boundary", 256, 2718)),
+    }, indent=2))