commit 47427245b268ca094495cda0d17f1bfd64c2b462 Author: dingbin.wang Date: Thu Oct 15 16:07:14 2026 +0800 jev v0.2 dataset bundle + training script diff --git a/README.md b/README.md new file mode 100644 index 0000000..211664f --- /dev/null +++ b/README.md @@ -0,0 +1,3 @@ +# jev-training-data + +jev_dataset_v0.2 bundle + train_jev.py diff --git a/bundle.tar.gz b/bundle.tar.gz new file mode 100644 index 0000000..57683fd Binary files /dev/null and b/bundle.tar.gz differ diff --git a/train_jev.py b/train_jev.py new file mode 100644 index 0000000..fc3bb55 --- /dev/null +++ b/train_jev.py @@ -0,0 +1,124 @@ +#!/usr/bin/env python3 +"""train_jev.py — jev v1 训练: xlm-roberta-base 判别式文档级相关性打分(校准概率) +输入: --data 目录下 train.jsonl / calibration.jsonl / eval.jsonl / injection.jsonl +输出: --out 目录下 model/(最佳权重) + metrics.json + train.log 由外层 tee 落盘 +损失: BCE(hard) + 0.5*MSE(sigmoid(logit), soft) — 校准概率是 jev 的第一公民 +指标: AUC + ECE(15 bins) @ calibration, 每 epoch 选最佳 +""" +import json, os, argparse, math, random +import torch +import torch.nn.functional as F +from torch.utils.data import Dataset, DataLoader +from transformers import AutoTokenizer, AutoModelForSequenceClassification, get_cosine_schedule_with_warmup + +def load_jsonl(p): + return [json.loads(l) for l in open(p, encoding="utf-8")] + +class PairDS(Dataset): + def __init__(self, rows, tok, max_len=512): + self.rows, self.tok, self.max_len = rows, tok, max_len + def __len__(self): + return len(self.rows) + def __getitem__(self, i): + r = self.rows[i] + enc = self.tok(r["query"], r["doc_text"][:2200], truncation="only_second", + max_length=self.max_len, padding="max_length", return_tensors="pt") + soft = r.get("soft_label") + soft = float(soft) if soft is not None else float(r["hard_label"]) + return {"input_ids": enc["input_ids"][0], "attention_mask": enc["attention_mask"][0], + "hard": torch.tensor(float(r["hard_label"])), "soft": torch.tensor(soft)} + +def ece(probs, labels, n_bins=15): + bins = [[] for _ in range(n_bins)] + for p, y in zip(probs, labels): + bins[min(int(p * n_bins), n_bins - 1)].append((p, y)) + e = 0.0 + for b in bins: + if not b: + continue + conf = sum(p for p, _ in b) / len(b) + acc = sum(y for _, y in b) / len(b) + e += len(b) / len(probs) * abs(conf - acc) + return e + +def auc(probs, labels): + pos = sorted(p for p, y in zip(probs, labels) if y > 0.5) + neg = sorted(p for p, y in zip(probs, labels) if y <= 0.5) + if not pos or not neg: + return 0.5 + import bisect + wins = sum(bisect.bisect_left(neg, p) for p in pos) + return wins / (len(pos) * len(neg)) + +def evaluate(model, loader, dev): + model.eval() + probs, labels = [], [] + with torch.no_grad(): + for b in loader: + out = model(input_ids=b["input_ids"].to(dev), attention_mask=b["attention_mask"].to(dev)) + probs += torch.sigmoid(out.logits[:, 0]).tolist() + labels += b["hard"].tolist() + model.train() + return {"auc": round(auc(probs, labels), 4), "ece": round(ece(probs, labels), 4), + "n": len(labels)} + +def main(): + ap = argparse.ArgumentParser() + ap.add_argument("--data", required=True) + ap.add_argument("--out", required=True) + ap.add_argument("--base", default="xlm-roberta-base") + ap.add_argument("--epochs", type=int, default=3) + ap.add_argument("--bs", type=int, default=32) + ap.add_argument("--lr", type=float, default=2e-5) + ap.add_argument("--lambda-soft", type=float, default=0.5) + args = ap.parse_args() + os.makedirs(args.out, exist_ok=True) + + dev = "cuda" if torch.cuda.is_available() else "cpu" + print(f"[jev] device={dev} base={args.base}", flush=True) + tok = AutoTokenizer.from_pretrained(args.base) + model = AutoModelForSequenceClassification.from_pretrained(args.base, num_labels=1).to(dev) + + train = [r for r in load_jsonl(os.path.join(args.data, "train.jsonl")) if r.get("hard_label") in (0, 1)] + calib = load_jsonl(os.path.join(args.data, "calibration.jsonl")) + print(f"[jev] train={len(train)} calib={len(calib)}", flush=True) + + dl = DataLoader(PairDS(train, tok), batch_size=args.bs, shuffle=True, num_workers=2, drop_last=True) + dl_cal = DataLoader(PairDS(calib, tok), batch_size=64, num_workers=2) + + steps = len(dl) * args.epochs + opt = torch.optim.AdamW(model.parameters(), lr=args.lr, weight_decay=0.01) + sch = get_cosine_schedule_with_warmup(opt, int(0.06 * steps), steps) + scaler = torch.amp.GradScaler(enabled=dev == "cuda") + best = {"auc": 0.0} + gstep = 0 + for ep in range(args.epochs): + for b in dl: + with torch.autocast("cuda", enabled=dev == "cuda"): + out = model(input_ids=b["input_ids"].to(dev), attention_mask=b["attention_mask"].to(dev)) + logits = out.logits[:, 0].float() + hard = b["hard"].to(dev); soft = b["soft"].to(dev) + loss = F.binary_cross_entropy_with_logits(logits, hard) + \ + args.lambda_soft * F.mse_loss(torch.sigmoid(logits), soft) + opt.zero_grad() + scaler.scale(loss).backward() + scaler.unscale_(opt) + torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0) + scaler.step(opt); scaler.update(); sch.step() + gstep += 1 + if gstep % 100 == 0: + print(f"[jev] ep{ep} step{gstep}/{steps} loss={loss.item():.4f}", flush=True) + m = evaluate(model, dl_cal, dev) + print(f"[jev] epoch {ep} calibration: {m}", flush=True) + if m["auc"] > best["auc"]: + best = {"auc": m["auc"], "ece": m["ece"], "epoch": ep} + model.save_pretrained(os.path.join(args.out, "model")) + tok.save_pretrained(os.path.join(args.out, "model")) + json.dump({"best": best, "base": args.base, "train_pairs": len(train)}, + open(os.path.join(args.out, "metrics.json"), "w"), indent=1) + print(f"[jev] DONE best={best}", flush=True) + with open("/dev/termination-log", "w") as f: + f.write(f"jev training done best={best}") + +if __name__ == "__main__": + main()