85 lines
3.9 KiB
Python
Raw Permalink Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

#!/usr/bin/env python3
"""14_build_v04.py — v0.4: v0.3 train + 泛化页负例(×3) ; eval_v3 = eval_v2 + 标签修正
输出: data/jev_dataset_v0.4/ + data/raw/parts_v04/ + eval_v3.jsonl
"""
import json, os, sys, hashlib, shutil, tarfile
HERE = os.path.dirname(os.path.abspath(__file__))
D = lambda n: os.path.join(HERE, "..", "data", n)
def main():
upweight = int(sys.argv[1]) if len(sys.argv) > 1 else 3
corpus = {json.loads(l)["doc_id"]: json.loads(l) for l in open(D("corpus.jsonl"), encoding="utf-8")}
# 1. 标签修正 → eval_v3
fixes = json.load(open(D("eval_label_fixes.json"), encoding="utf-8"))
fix_by_pair = {f["pair_id"]: f for f in fixes}
eval_rows = [json.loads(l) for l in open(D("eval_v2.jsonl"), encoding="utf-8")]
n_flip = 0
eval_v3 = []
for r in eval_rows:
f = fix_by_pair.get(r["pair_id"])
if f and f["verdict"] == "flip_to_positive":
r = dict(r)
r["hard_label"], r["grade"] = 1, "full"
r["soft_label"] = round(sum(j["p"] for j in f["judges"] if j) / 2, 3)
r["review"]["iter3"] = "flip_to_positive"
n_flip += 1
eval_v3.append(r)
with open(D("eval_v3.jsonl"), "w", encoding="utf-8") as f:
for r in eval_v3:
f.write(json.dumps(r, ensure_ascii=False) + "\n")
print(f"eval_v3: {len(eval_v3)} 对 (修正翻转 {n_flip})")
# 2. train = v0.3 train + 泛化负例 ×3
train = [json.loads(l) for l in open(D("jev_dataset_v0.3/train.jsonl"), encoding="utf-8")]
negs = [json.loads(l) for l in open(D("generic_page_negatives.jsonl"), encoding="utf-8")]
added = 0
for r in negs:
for w in range(upweight):
train.append({
"pair_id": hashlib.sha256((r["query"] + "\x00" + r["doc_id"] + f"\x00g{w}").encode()).hexdigest()[:16],
"query": r["query"], "query_type": "generic_page_negative",
"doc_id": r["doc_id"], "doc_text": corpus[r["doc_id"]]["text"],
"doc_truncated": corpus[r["doc_id"]]["truncated"],
"hard_label": 0, "grade": "none", "soft_label": r["soft_label"],
"judge_scores": [j["p"] for j in r["judges"] if j],
"difficulty": "generic_page", "mutation": None, "split": "train",
"provenance": {**r["provenance"], "upweight": w}})
added += 1
outdir = D("jev_dataset_v0.4")
if os.path.exists(outdir):
shutil.rmtree(outdir)
os.makedirs(outdir)
with open(os.path.join(outdir, "train.jsonl"), "w", encoding="utf-8") as f:
for r in train:
f.write(json.dumps(r, ensure_ascii=False) + "\n")
for name in ["calibration", "injection"]:
shutil.copy(D(f"jev_dataset_v0.3/{name}.jsonl"), os.path.join(outdir, name + ".jsonl"))
shutil.copy(D("eval_v3.jsonl"), os.path.join(outdir, "eval.jsonl"))
json.dump({"tag": "v0.4", "train": len(train), "generic_neg_added": added, "upweight": upweight,
"eval_label_fixed": n_flip},
open(os.path.join(outdir, "manifest_delta.json"), "w"), ensure_ascii=False, indent=1)
print(f"v0.4: train {len(train)} 对 (泛化负例 {added}, ×{upweight})")
# 3. 打包分片
parts_dir = D("raw/parts_v04")
shutil.rmtree(parts_dir, ignore_errors=True)
os.makedirs(parts_dir)
bundle = D("raw/jev-v0.4-bundle.tar.gz")
with tarfile.open(bundle, "w:gz") as tf:
for name in ["train.jsonl", "calibration.jsonl", "eval.jsonl", "injection.jsonl", "manifest_delta.json"]:
tf.add(os.path.join(outdir, name), arcname=name)
i = 0
with open(bundle, "rb") as f:
while True:
b = f.read(5 * 1024 * 1024)
if not b:
break
open(os.path.join(parts_dir, f"jev_part_{i:02d}"), "wb").write(b)
i += 1
print(f"bundle → {i} 分片, 目录 {os.path.normpath(parts_dir)}")
if __name__ == "__main__":
sys.exit(main())