193 lines
8.5 KiB
Python

#!/usr/bin/env python3
"""train_jev_v2.py — jev v2 训练: 多底座判别式打分(校准概率), 支持 modernbert/xlm-roberta
用法同 train_jev.py; --base 指向任意 encoder 底座, 架构与 LoRA 目标层自动探测
"""
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
from peft import LoraConfig, get_peft_model
def lora_targets(model):
"""按底座架构自动探测 LoRA 目标层"""
names = {n for n, _ in model.named_modules()}
if any(n.endswith("Wqkv") for n in names):
return ["Wqkv", "Wo"] # ModernBERT 融合注意力
if any(n.endswith("query") for n in names):
return ["query", "key", "value"] # XLM-R / BERT 系
if any(n.endswith("q_proj") for n in names):
return ["q_proj", "k_proj", "v_proj"] # 通用命名
raise RuntimeError("未识别的注意力层命名, 请手动指定 --lora-targets")
def load_jsonl(p):
return [json.loads(l) for l in open(p, encoding="utf-8")]
def self_save_live(model_dir):
"""每个最佳 epoch 后把模型切 18MB 分片写进 git 仓库目录, 由监控循环自动推送(防僵死/抢占丢权重)"""
import shutil
live = os.path.join(os.environ.get("JEV_REPO", "/tmp/jd"), "model_live")
shutil.rmtree(live, ignore_errors=True)
os.makedirs(live, exist_ok=True)
for fn in sorted(os.listdir(model_dir)):
p = os.path.join(model_dir, fn)
if os.path.getsize(p) <= 18 * 1024 * 1024:
shutil.copy(p, os.path.join(live, fn))
else:
with open(p, "rb") as f:
i = 0
while True:
b = f.read(18 * 1024 * 1024)
if not b:
break
open(os.path.join(live, f"{fn}.part{i:03d}"), "wb").write(b)
i += 1
print(f"[jev] live-checkpoint 已写 {live}", flush=True)
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]
# longest_first 截断(全 tokenizers 版本通用): query 短不会受损, doc 预截到 1200 字符
enc = self.tok(r["query"], r["doc_text"][:1200], truncation=True,
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)
ap.add_argument("--lora-targets", default=None, help="逗号分隔; 缺省自动探测")
args = ap.parse_args()
os.makedirs(args.out, exist_ok=True)
dev = "cuda" if torch.cuda.is_available() else "cpu"
try:
import tokenizers
tok_ver = tokenizers.__version__
except Exception:
tok_ver = "?"
print(f"[jev] device={dev} base={args.base} tokenizers={tok_ver}", flush=True)
try:
tok = AutoTokenizer.from_pretrained(args.base)
model = AutoModelForSequenceClassification.from_pretrained(args.base, num_labels=1).to(dev)
targets = args.lora_targets.split(",") if args.lora_targets else lora_targets(model)
print(f"[jev] LoRA targets: {targets}", flush=True)
lora = LoraConfig(r=16, lora_alpha=32, lora_dropout=0.05,
target_modules=targets, task_type="SEQ_CLS")
model = get_peft_model(model, lora)
model.print_trainable_parameters()
except Exception:
import traceback
tail = traceback.format_exc()[-1200:]
try:
with open("/dev/termination-log", "a") as f:
f.write("\nEXC-TAIL:\n" + tail)
except Exception:
pass
raise
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)
# 金丝雀: 双语言切片逐 epoch 评估, best 按 min(auc_en, auc_zh) 选择(防单语言偏科)
canaries = {}
for lang in ["en", "zh"]:
p = os.path.join(args.data, f"canary_{lang}.jsonl")
if os.path.exists(p):
dl_c = DataLoader(PairDS(load_jsonl(p), tok), batch_size=64, num_workers=2)
canaries[lang] = dl_c
print(f"[jev] canaries: {list(canaries) or '无(单语言模式)'}", flush=True)
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)
sel_score = m["auc"]
for lang, dl_c in canaries.items():
cm = evaluate(model, dl_c, dev)
print(f"[jev] epoch {ep} canary_{lang}: {cm}", flush=True)
sel_score = min(sel_score, cm["auc"]) # 最差语言决定选择
if sel_score > 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"))
self_save_live(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()