#!/usr/bin/env python3 """train_laya.py — Laya 官方 finetune() 单卡封装 + 内置 eval_v2 评分 用法: python3 train_laya.py <模型快照目录> <输出目录> """ import json, os, sys data_path = sys.argv[1] if len(sys.argv) > 1 else "/tmp/jd/laya_v05_train_2400.jsonl" model_dir = sys.argv[2] if len(sys.argv) > 2 else None output_dir = sys.argv[3] if len(sys.argv) > 3 else "/tmp/laya_out" eval_path = "/tmp/jd/eval_v2.jsonl" from laya.train import TrainConfig, finetune cfg = TrainConfig( freeze_encoder=True, # 只训决策头: 从构造上杜绝 encoder 灾难性遗忘 epochs=2, micro_batch=8, grad_accum=8, amp=True, log_every=50, ) summary = finetune(data_path, model_dir, output_dir, cfg) print("[laya-ft] finetune summary:", json.dumps(summary, ensure_ascii=False)[:400], flush=True) # ---- eval_v2 评分 ---- import laya agent = laya.Agent(output_dir, device="cuda") rows = [json.loads(l) for l in open(eval_path, encoding="utf-8")] INSTR = "Does this document contain the information needed to answer the user's question? Answer based only on the document content." probs, labels = [], [] for i, r in enumerate(rows): state = f"用户问题: {r['query']}\n\n文档内容:\n{r['doc_text'][:3000]}" try: res = agent.predict(state, {"relevant": {"type": "noul", "instructions": INSTR}}) p = float(res["answers"]["relevant"]["noul"]) except Exception: p = 0.5 probs.append(p) labels.append(r["hard_label"]) if (i + 1) % 25 == 0: print(f"[eval] {i+1}/{len(rows)}", flush=True) import bisect pos = sorted(p for p, y in zip(probs, labels) if y == 1) neg = sorted(p for p, y in zip(probs, labels) if y == 0) auc = sum(bisect.bisect_left(neg, p) for p in pos) / (len(pos) * len(neg)) if pos and neg else 0.0 low = sum(1 for p in pos if p < 0.5) fp = sum(1 for p in neg if p >= 0.7) print(f"[laya-ft] eval_v2: AUC={auc:.4f} pos<0.5={low}/22 neg>=0.7={fp}/85", flush=True) with open("/dev/termination-log", "w") as f: f.write(f"LAYA-FT AUC={auc:.4f} pos<0.5={low}/22 neg>=0.7={fp}/85") print("LAYA-FT-DONE", flush=True)