v0.5-nordk: RDK全剥离纯公开数据 + Laya noul 训练脚本

This commit is contained in:
dingbin.wang 2026-10-21 19:10:56 +08:00
parent 700afc572a
commit db28cdb6f0
5 changed files with 184 additions and 0 deletions

184
train_laya_noul.py Normal file
View File

@ -0,0 +1,184 @@
#!/usr/bin/env python3
"""train_laya_noul.py — Laya DecisionModel 自研训练循环 (noul 二分类)
绕开未发布的 laya.train 模块, 直接用 build_sequence + DecisionModel.forward
用法: python3 train_laya_noul.py <mmBERT目录> <输出目录>
数据: /tmp/jd/{train,calibration,canary_en,canary_zh}.jsonl
"""
import json, os, sys, random, bisect
import torch
import torch.nn.functional as F
from torch.utils.data import Dataset, DataLoader
from transformers import AutoTokenizer, AutoConfig, AutoModel
from peft import LoraConfig, get_peft_model
sys.path.insert(0, "/tmp/layasrc")
from laya.common import build_sequence, build_model
# noul: markers = [false, true], true 位置 idx=1
def load_jsonl(p):
return [json.loads(l) for l in open(p, encoding="utf-8")]
class NoulDS(Dataset):
def __init__(self, rows, tok, max_len=512):
self.rows, self.tok, self.max_len = rows, tok, max_len
self.q = {"t": "noul", "ins": INSTR, "crit": None}
def __len__(self):
return len(self.rows)
def __getitem__(self, i):
r = self.rows[i]
state = f"用户问题: {r['query']}\n\n文档内容:\n{r['doc_text'][:2500]}"
ids, markers = build_sequence(self.tok, state, self.q, self.max_len, 192)
# markers[-1] 是 true marker (noul: [false, true])
true_pos = markers[-1]
# pad
pad = self.max_len - len(ids)
input_ids = ids + [self.tok.pad_token_id or 0] * pad
attn = [1] * len(ids) + [0] * pad
return {"input_ids": torch.tensor(input_ids), "attention_mask": torch.tensor(attn),
"marker_pos": torch.tensor(markers), "marker_mask": torch.tensor([1] * len(markers) + [0] * (8 - len(markers))),
"qtype": torch.tensor(2),
"label": torch.tensor(float(r["hard_label"])),
"soft": torch.tensor(float(r.get("soft_label") or r["hard_label"]))}
INSTR = "Does this document contain the information needed to answer the user's question? Answer based only on the document content."
def collate(batch):
keys = batch[0].keys()
return {k: torch.stack([b[k] for b in batch]) for k in keys}
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 evaluate(model, loader, dev):
model.eval()
probs, labels = [], []
with torch.no_grad():
for b in loader:
logits, _ = model(input_ids=b["input_ids"].to(dev), attention_mask=b["attention_mask"].to(dev),
marker_pos=b["marker_pos"].to(dev), marker_mask=b["marker_mask"].to(dev),
qtype=b["qtype"].to(dev))
# noul: markers = [false, true], true marker 是最后一个有效 marker
# logits shape: (batch, 8) — 取 true marker (index 1 in each pair... 简化: 取 max)
# 实际: logits[..., 1] 对应 true marker (noul 只有 2 markers)
p = torch.softmax(logits, -1)[..., 1] # index 1 = true
probs += p.tolist()
labels += b["label"].tolist()
model.train()
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
return {"auc": round(auc, 4), "ece": round(ece(probs, labels), 4), "n": len(labels)}
def main():
import argparse
ap = argparse.ArgumentParser()
ap.add_argument("--data", default="/tmp/jd")
ap.add_argument("--base", default="/tmp/base")
ap.add_argument("--out", default="/tmp/laya_out")
ap.add_argument("--epochs", type=int, default=2)
ap.add_argument("--bs", type=int, default=16)
ap.add_argument("--lr", type=float, default=1e-4)
ap.add_argument("--eval-data", default="/tmp/jd/eval_v2.jsonl")
args = ap.parse_args()
os.makedirs(args.out, exist_ok=True)
dev = "cuda" if torch.cuda.is_available() else "cpu"
from transformers import AutoTokenizer, AutoModelForSequenceClassification
# 加载 mmBERT bare encoder
enc_cfg = AutoConfig.from_pretrained(args.base)
enc = AutoModel.from_config(enc_cfg)
d = enc.config.hidden_size
# 构建 DecisionModel (从 laya.common 的架构)
import laya_common as lc
dm = lc.DecisionModel(enc, head_layers=2, n_act=2)
dm.to(dev)
# LoRA 只在 encoder attention
lora = LoraConfig(r=16, lora_alpha=32, lora_dropout=0.05,
target_modules=["Wqkv"], task_type=None)
# 手动对 encoder 加 LoRA (peft 对 bare nn.Module 用 get_peft_model)
dm.encoder = get_peft_model(dm.encoder, lora)
dm.print_trainable_parameters()
dm.train()
tok = AutoTokenizer.from_pretrained(args.base)
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)
ds = NoulDS(train, tok)
dl = DataLoader(ds, batch_size=args.bs, shuffle=True, num_workers=2, drop_last=True,
collate_fn=collate)
dl_cal = DataLoader(NoulDS(calib, tok), batch_size=32, collate_fn=collate)
canaries = {}
for lang in ["en", "zh"]:
p = os.path.join(args.data, f"canary_{lang}.jsonl")
if os.path.exists(p):
canaries[lang] = DataLoader(NoulDS(load_jsonl(p), tok), batch_size=32, collate_fn=collate)
n_params = sum(p.numel() for p in dm.parameters() if p.requires_grad)
steps = len(dl) * args.epochs
opt = torch.optim.AdamW([p for p in dm.parameters() if p.requires_grad], lr=args.lr, weight_decay=0.01)
sch = get_cosine_schedule_with_warmup(opt, int(0.06 * steps), steps)
best = {"sel": 0.0}
gstep = 0
for ep in range(args.epochs):
for b in dl:
logits, act = dm(input_ids=b["input_ids"].to(dev), attention_mask=b["attention_mask"].to(dev),
marker_pos=b["marker_pos"].to(dev), marker_mask=b["marker_mask"].to(dev),
qtype=b["qtype"].to(dev))
# noul markers = [false, true]: true marker 是每对中最后一个有效
# marker_mask (batch, 8): 每样本 2 个有效 marker
# 简化: 取 logits[..., 1] (true marker 位置)
true_logit = logits[:, 1].float()
hard = b["label"].to(dev); soft = b["soft"].to(dev)
loss = F.binary_cross_entropy_with_logits(true_logit, hard) + \
0.5 * F.mse_loss(torch.sigmoid(true_logit), soft)
opt.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm_([p for p in dm.parameters() if p.requires_grad], 1.0)
opt.step(); sch.step()
gstep += 1
if gstep % 100 == 0:
print(f"[jev] ep{ep} step{gstep}/{steps} loss={loss.item():.4f}", flush=True)
# calibration + canaries 评估
dm.eval()
def _auc(loader):
ps, ls = [], []
with torch.no_grad():
for b in loader:
lg, _ = dm(input_ids=b["input_ids"].to(dev), attention_mask=b["attention_mask"].to(dev),
marker_pos=b["marker_pos"].to(dev), marker_mask=b["marker_mask"].to(dev),
qtype=b["qtype"].to(dev))
p = torch.softmax(lg.float(), -1)[..., 1]
ps += p.tolist(); ls += b["label"].tolist()
pos = sorted(p for p, y in zip(ps, ls) if y == 1)
neg = sorted(p for p, y in zip(ps, ls) if y == 0)
return sum(bisect.bisect_left(neg, p) for p in pos) / (len(pos) * len(neg)) if pos and neg else 0.0
auc_cal = _auc(dl_cal)
msg = f"[jev] epoch {ep} calib_auc={auc_cal:.4f}"
sel = auc_cal
for lang, dl_c in canaries.items():
a = _auc(dl_c)
msg += f" canary_{lang}={a:.4f}"
sel = min(sel, a)
print(msg, flush=True)
if sel > best["sel"]:
best = {"sel": sel, "calib_auc": auc_cal, "epoch": ep}
torch.save(dm.state_dict(), os.path.join(args.out, "model.pt"))
dm.train()
print(f"[jev] DONE best={best}", flush=True)
with open("/dev/termination-log", "w") as f:
f.write(f"LAYA-NOUL DONE best={best}")
if __name__ == "__main__":
main()

BIN
v05n/jev_part_00 Normal file

Binary file not shown.

BIN
v05n/jev_part_01 Normal file

Binary file not shown.

BIN
v05n/jev_part_02 Normal file

Binary file not shown.

BIN
v05n/jev_part_03 Normal file

Binary file not shown.