train_all r40: clean-soft MSE mask + random-window aug + max_len 1024 + temperature calibration

This commit is contained in:
dingbin.wang 2026-11-06 00:48:33 +08:00
parent 1bc40935f6
commit d345e25db4

View File

@ -40,6 +40,13 @@ def ece(probs, labels, nbins=15):
sum(p for p, _ in b) / len(b) - sum(y for _, y in b) / len(b))
return e
def bisect_auc(probs, labels):
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)
if not pos or not neg:
return 0.0
return sum(bisect.bisect_left(neg, p) for p in pos) / (len(pos) * len(neg))
def phase_data():
"""v05n tar.gz 分片重组 → DATA_DIR/train.jsonl"""
import glob, shutil, tarfile
@ -169,6 +176,7 @@ def run(args):
MMBERT_DIR = args.base
if args.out:
OUT_DIR = args.out
os.makedirs(OUT_DIR, exist_ok=True) # 让早期 log 能落盘
if not phase_env():
sys.exit(1)
@ -177,6 +185,8 @@ def run(args):
if not (args.tiny or phase_mmbert()):
sys.exit(1)
import random
import math
import torch
import torch.nn.functional as F
from torch.utils.data import Dataset, DataLoader
@ -184,28 +194,50 @@ def run(args):
from laya.common import build_sequence
dev = "cuda" if torch.cuda.is_available() else "cpu"
log(f"device={dev} tiny={args.tiny}")
log(f"device={dev} tiny={args.tiny} max_len={args.max_len} head_len={args.head_len} "
f"win={args.win_tokens} aug_p={args.aug_p}")
dm, tok = build_model(dev, tiny=args.tiny)
class NoulDS(Dataset):
def __init__(self, rows, max_len=512, head_len=192, doc_cap=2500):
self.rows, self.max_len, self.head_len, self.doc_cap = rows, max_len, head_len, doc_cap
def __init__(self, rows, max_len=None, head_len=None, train_aug=False):
self.rows = rows
self.max_len = max_len or args.max_len
self.head_len = head_len or args.head_len
self.win_tokens = args.win_tokens
self.aug_p = args.aug_p
self.train_aug = train_aug
self.q = {"t": "noul", "ins": INSTR, "crit": None}
self.rng = random.Random(20261105)
def _doc_window(self, text):
if not self.train_aug:
return text
ids = tok(text[:12000], add_special_tokens=False)["input_ids"]
if len(ids) <= self.win_tokens:
return text
if self.rng.random() < self.aug_p:
s = self.rng.randrange(0, len(ids) - self.win_tokens + 1)
return tok.decode(ids[s:s + self.win_tokens])
return text
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'][:self.doc_cap]}"
state = f"用户问题: {r['query']}\n\n文档内容:\n{self._doc_window(r['doc_text'])}"
ids, markers = build_sequence(tok, state, self.q, self.max_len, self.head_len)
pad = self.max_len - len(ids)
mm = [True] * len(markers) + [False] * (8 - len(markers))
soft = r.get("soft_label")
if soft is None:
soft = r["hard_label"]
soft_clean = float(soft) if soft in (0, 1) else -1.0
return {"input_ids": torch.tensor(ids + [0] * pad),
"attention_mask": torch.tensor([1] * len(ids) + [0] * pad),
"marker_pos": torch.tensor(markers + [0] * (8 - len(markers))),
"marker_mask": torch.tensor(mm),
"qtype": torch.tensor(2),
"label": torch.tensor(float(r["hard_label"])),
"soft": torch.tensor(float(r.get("soft_label") or r["hard_label"]))}
"soft": torch.tensor(float(soft)),
"soft_clean": torch.tensor(soft_clean)}
def collate(batch):
return {k: torch.stack([b[k] for b in batch]) for k in batch[0]}
@ -246,7 +278,7 @@ def run(args):
os.makedirs(OUT_DIR, exist_ok=True)
nw = 0 if args.tiny else 2
dl = DataLoader(NoulDS(train), batch_size=args.bs, shuffle=True,
dl = DataLoader(NoulDS(train, train_aug=True), batch_size=args.bs, shuffle=True,
num_workers=nw, drop_last=True, collate_fn=collate)
dl_cal = DataLoader(NoulDS(calib), batch_size=32, collate_fn=collate)
@ -266,9 +298,13 @@ def run(args):
logits = fwd(b)
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)
# MSE 校准项只作用于 soft∈{0,1} 的干净行 (47% 正例 soft=0.5 是标注缺失
# 的填充值, 0.5*MSE 会把概率系统性拉向 0.5 → r39 校准差的总根因)
soft_clean = b["soft_clean"].to(dev)
mask = soft_clean >= 0
loss = F.binary_cross_entropy_with_logits(true_logit, hard)
if mask.any():
loss = loss + 0.2 * F.mse_loss(torch.sigmoid(true_logit)[mask], soft_clean[mask])
opt.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm_(
@ -286,13 +322,56 @@ def run(args):
log(f"saved best (auc={best_auc:.4f})")
log(f"TRAINING-DONE best_auc={best_auc:.4f}")
# ===== 温度校准: best 模型在 calib 全集上 grid-search T (min NLL) =====
calib_info = {"best_auc": best_auc}
best_path = os.path.join(OUT_DIR, "model.pt")
if os.path.exists(best_path):
dm.load_state_dict(torch.load(best_path, map_location=dev, weights_only=False))
was = dm.training
dm.eval()
z_all, y_all = [], []
with torch.no_grad():
for b in dl_cal:
lg = fwd(b, no_grad=True)
z_all += lg[..., 1].float().tolist()
y_all += b["label"].tolist()
if was:
dm.train()
def nll_at(T):
tot = 0.0
for z, y in zip(z_all, y_all):
p = 1.0 / (1.0 + math.exp(-z / T))
p = min(max(p, 1e-6), 1 - 1e-6)
tot -= math.log(p) if y == 1 else math.log(1 - p)
return tot / len(z_all)
probs_raw = [1 / (1 + math.exp(-z)) for z in z_all]
best_T, n_at_best = 1.0, nll_at(1.0)
T = 0.30
while T <= 4.001:
n = nll_at(T)
if n < n_at_best:
best_T, n_at_best = T, n
T += 0.05
probs_T = [1 / (1 + math.exp(-z / best_T)) for z in z_all]
calib_info = {
"best_auc": best_auc, "calib_n": len(z_all),
"auc_raw": round(bisect_auc(probs_raw, y_all), 4),
"ece_raw": round(ece(probs_raw, y_all), 4),
"temperature": round(best_T, 2),
"auc_calibrated": round(bisect_auc(probs_T, y_all), 4),
"ece_calibrated": round(ece(probs_T, y_all), 4),
}
log(f"CALIBRATION T={best_T:.2f} raw(ece={calib_info['ece_raw']}) "
f"calibrated(ece={calib_info['ece_calibrated']}) auc={calib_info['auc_calibrated']}")
json.dump({"best_auc": best_auc, "train_pairs": len(train),
"steps": gstep, "epochs": args.epochs,
"pretrain_calib": m0, "final_calib": m},
"pretrain_calib": m0, "final_calib": m, "calibration": calib_info},
open(os.path.join(OUT_DIR, "training_log.json"), "w"),
ensure_ascii=False, indent=1)
os.makedirs(os.path.join(DATA_DIR, "logs"), exist_ok=True)
json.dump({"best_auc": best_auc, "train_pairs": len(train), "steps": gstep},
json.dump({"best_auc": best_auc, "train_pairs": len(train), "steps": gstep,
"calibration": calib_info},
open(os.path.join(DATA_DIR, "logs", f"{args.tag}_result.json"), "w"),
ensure_ascii=False, indent=1)
if not args.tiny:
@ -370,4 +449,8 @@ if __name__ == "__main__":
ap.add_argument("--out", default=None, help="覆盖 OUT_DIR")
ap.add_argument("--tiny", action="store_true", help="冒烟模式: 随机初始化 + 64 条数据")
ap.add_argument("--tag", default="r36", help="产物回传标签")
ap.add_argument("--max_len", type=int, default=512, help="序列总长 (r40=1024)")
ap.add_argument("--head_len", type=int, default=192, help="question 头部预算")
ap.add_argument("--win_tokens", type=int, default=800, help="随机窗口 token 数")
ap.add_argument("--aug_p", type=float, default=0.4, help="随机窗口采样概率")
run(ap.parse_args())