#!/usr/bin/env python3 # data_contract_check.py (v3) -- train-file "data contract" preflight (数据契约检查). CPU only, read only. # Motivation: wave E (2026-10-30) trained ~2 h x 4 GPUs on 今译 rows whose user turn carried the task wrapper # twice; no existing preflight looked at row CONTENT FORMAT. This check fails that file in seconds. # # The contract is LEARNED from a reference file (default stage2_v1 = the format the eval sets were built from): # template markers T[task] = 【…】 markers present in >= 90% of reference rows of that task # (content labels such as slip numbers 【98】【志2】【4+5】 are therefore never treated as wrappers). # Checks, per row of the candidate file: # C1 wrapper multiplicity: a template marker occurs more often than its max multiplicity in the reference. # C2 template parity: the sequence of template markers is not one the reference uses for that task. # C3 role shape: roles must be exactly system,user,assistant with a non-empty assistant turn. # C0 unknown task: a system-prompt prefix that the reference never uses (reported; fails the contract). # report-only: Han/Qin vs Chu cue counts, so "楚简" labels can be sanity-checked. # History: v1/v2 counted every 【…】 marker and false-failed the reference itself on slip labels (fixed in v3). # Usage: data_contract_check.py TRAIN.jsonl [REFERENCE.jsonl] -> exit 0 PASS / 1 FAIL import collections, json, re, sys MARK = re.compile(r"【[^】]{1,16}】") HAN = re.compile(r"肩水|居延|金关|悬泉|敦煌|候长|燧长|隧长|啬夫|里耶|迁陵|甲渠|鄣卒|戍卒|候官|令史|亭长") CHU = re.compile(r"楚|郢|[戠𢍰]|\{[^}]+\}") TEMPLATE_FRAC = 0.90 def rows(p): return [json.loads(l) for l in open(p, encoding="utf-8") if l.strip()] def task_of(r): s = [m["content"] for m in r["messages"] if m["role"] == "system"] return s[0][:24] if s else "" def user_of(r): u = [m["content"] for m in r["messages"] if m["role"] == "user"] return u[0] if u else "" def learn(R): by_task = collections.defaultdict(list) for r in R: by_task[task_of(r)].append(MARK.findall(user_of(r))) T, maxmult, skeletons = {}, {}, {} for t, lists in by_task.items(): n = len(lists) present = collections.Counter(m for ms in lists for m in set(ms)) T[t] = {m for m, c in present.items() if c / n >= TEMPLATE_FRAC} maxmult[t] = {m: max(ms.count(m) for ms in lists) for m in T[t]} skeletons[t] = {tuple(m for m in ms if m in T[t]) for ms in lists} return T, maxmult, skeletons def main(): train = sys.argv[1] ref = sys.argv[2] if len(sys.argv) > 2 else "/root/autodl-tmp/llm_rag_20261023/data/stage2_v1.jsonl" R, X = rows(ref), rows(train) T, maxmult, skeletons = learn(R) viol, examples = collections.Counter(), {} per_task = collections.defaultdict(collections.Counter) def hit(kind, t, i): viol[kind] += 1; per_task[t][kind] += 1; examples.setdefault(kind, i) for i, r in enumerate(X): t, u = task_of(r), user_of(r) per_task[t]["rows"] += 1 per_task[t]["han_cue"] += bool(HAN.search(u)) per_task[t]["chu_cue"] += bool(CHU.search(u)) if [m["role"] for m in r["messages"]] != ["system", "user", "assistant"] or not r["messages"][-1]["content"].strip(): hit("C3_role_shape", t, i) if t not in T: hit("C0_unknown_task", t, i); continue ms = MARK.findall(u) if any(ms.count(m) > maxmult[t][m] for m in T[t]): hit("C1_wrapper_multiplicity", t, i) if tuple(m for m in ms if m in T[t]) not in skeletons[t]: hit("C2_template_parity", t, i) print("train:", train, "rows", len(X), "| reference:", ref) for t in sorted(T): print(" template[%s] = %s" % (t[:14], sorted(T[t]))) for t, c in per_task.items(): print(" task %-26s rows=%5d C1=%5d C2=%5d C0=%4d han_cue=%4d chu_cue=%4d" % ( t, c["rows"], c["C1_wrapper_multiplicity"], c["C2_template_parity"], c["C0_unknown_task"], c["han_cue"], c["chu_cue"])) for k, i in examples.items(): print(" first %s at row %d: %s" % (k, i, json.dumps(user_of(X[i])[:110], ensure_ascii=False))) ok = not viol print("DATA_CONTRACT", "PASS" if ok else "FAIL", dict(viol)) sys.exit(0 if ok else 1) if __name__ == "__main__": main()