train_all r41: crosspair long negs (len×label balance) + agg-level T_agg calibration

This commit is contained in:
dingbin.wang 2026-11-06 10:22:13 +08:00
parent 27fa0f5c06
commit 276e20bb32

View File

@ -3,7 +3,7 @@
在 fat 容器内运行。数据和 laya_repo 模块通过 gitea clone 获取 (/tmp/jd)。
用法: python3 train_all.py [--epochs N] [--lr LR] [--bs N]
"""
import json, os, sys, bisect
import json, os, sys, bisect, random
INSTR = "Does this document contain the information needed to answer the user's question? Answer based only on the document content."
DATA_DIR = "/tmp/jd"
@ -47,6 +47,34 @@ def bisect_auc(probs, labels):
return 0.0
return sum(bisect.bisect_left(neg, p) for p in pos) / (len(pos) * len(neg))
def _lang(s):
return "zh" if any("一" <= c <= "鿿" for c in s) else "en"
def augment_crosspair_negs(rows, win_tokens, max_new=2500):
"""长文档 × 随机同语言其他 query → 长负例。
训练负例几乎都是短文档而 eval_v2 负例全是长文档——模型没学过
『长而无关 → 每窗都低分』。跨配对直接补齐 长度×标签 的失衡象限。"""
rng = random.Random(20261106)
nch_thr = {"zh": win_tokens, "en": win_tokens * 3} # 字符≈token 粗估
cand = [i for i, r in enumerate(rows)
if len(r.get("doc_text") or "") > nch_thr[_lang(r.get("doc_text") or "")]]
rng.shuffle(cand)
pools = {}
for i, r in enumerate(rows):
pools.setdefault(_lang(r["query"]), []).append(i)
added = []
for i in cand[:max_new]:
L = _lang(rows[i].get("doc_text") or "")
pool = pools[L]
j = rng.choice(pool)
if j == i:
continue
added.append({"pair_id": f"xp-{rows[i].get('pair_id', i)}",
"query": rows[j]["query"],
"doc_text": rows[i]["doc_text"],
"hard_label": 0, "soft_label": 0.0})
return added
def phase_data():
"""v05n tar.gz 分片重组 → DATA_DIR/train.jsonl"""
import glob, shutil, tarfile
@ -274,6 +302,10 @@ def run(args):
if args.tiny:
train, calib = train[:64], calib[:16]
log(" TINY mode: train[:64] calib[:16]")
else:
xp = augment_crosspair_negs(train, args.win_tokens, max_new=2500)
train = train + xp
log(f"crosspair long negs: +{len(xp)} (total {len(train)})")
log(f"train={len(train)} calib={len(calib)}")
os.makedirs(OUT_DIR, exist_ok=True)
@ -364,6 +396,90 @@ def run(args):
}
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']}")
# ===== 聚合后温度 T_agg: 合成 4-doc 长文档跑滑窗 max (生产聚合口径) =====
def agg_calib():
rng = random.Random(7)
pos_rows = [r for r in calib if r.get("hard_label") == 1][:150]
neg_rows = [r for r in calib if r.get("hard_label") == 0][:150]
neg_pool = [r for r in calib if r.get("hard_label") == 0]
if not pos_rows or not neg_rows:
return None
def windows_of(row):
docs = [row["doc_text"]] + [o["doc_text"] for o in rng.sample(neg_pool, 3)]
rng.shuffle(docs)
return [d[:4000] for d in docs]
zs, ys = [], []
B = 8
buf_ids, buf_lab, buf_idx = [], [], []
def flush_buf():
if not buf_ids:
return
b = {
"input_ids": torch.tensor([x[0] for x in buf_ids]).to(dev),
"attention_mask": torch.tensor(
[[1] * x[2] + [0] * (args.max_len - x[2]) for x in buf_ids]).to(dev),
"marker_pos": torch.tensor([x[1] for x in buf_ids]).to(dev),
"marker_mask": torch.tensor(
[[True] * x[3] + [False] * (8 - x[3]) for x in buf_ids]).bool().to(dev),
"qtype": torch.tensor([2] * len(buf_ids)).to(dev),
}
with torch.no_grad():
with torch.autocast("cuda", enabled=(dev == "cuda")):
lg, _ = dm(input_ids=b["input_ids"],
attention_mask=b["attention_mask"],
marker_pos=b["marker_pos"],
marker_mask=b["marker_mask"],
qtype=b["qtype"])
p1 = torch.softmax(lg.float(), -1)[..., 1].tolist()
for k in range(len(buf_ids)):
zs[buf_idx[k]][1].append(p1[k])
buf_ids.clear(); buf_lab.clear(); buf_idx.clear()
# zs: list of [label, [p_win...]]
zs = []
for r in pos_rows + neg_rows:
y = float(r["hard_label"])
row_pos = len(zs)
zs.append([y, []])
for d in windows_of(r):
state = f"用户问题: {r['query']}\n\n文档内容:\n{d}"
ids, markers = build_sequence(tok, state, {"t": "noul", "ins": INSTR, "crit": None},
args.max_len, args.head_len)
ids = ids + [0] * (args.max_len - len(ids))
buf_ids.append((ids, (markers + [0] * (8 - len(markers)))[:8],
len(ids), len(markers)))
buf_lab.append(y)
buf_idx.append(row_pos)
if len(buf_ids) >= B:
flush_buf()
flush_buf()
max_p = [max(ps) if ps else 0.5 for _, ps in zs]
labels = [y for y, _ in zs]
z_agg = [math.log(max(p, 1e-6) / max(1 - p, 1e-6)) for p in max_p]
def nll_z(zlist, T):
tot = 0.0
for z, y in zip(zlist, labels):
p = 1 / (1 + 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(zlist)
t_best, n_best = 1.0, nll_z(z_agg, 1.0)
T = 0.30
while T <= 6.001:
n = nll_z(z_agg, T)
if n < n_best:
t_best, n_best = T, n
T += 0.05
p_cal = [1 / (1 + math.exp(-z / t_best)) for z in z_agg]
return {"n": len(labels),
"auc_max_raw": round(bisect_auc(max_p, labels), 4),
"ece_max_raw": round(ece(max_p, labels), 4),
"T_agg": round(t_best, 2),
"ece_max_calibrated": round(ece(p_cal, labels), 4)}
agg = agg_calib()
if agg:
calib_info["agg"] = agg
log(f"AGG-CALIB {agg}")
json.dump({"best_auc": best_auc, "train_pairs": len(train),
"steps": gstep, "epochs": args.epochs,
"pretrain_calib": m0, "final_calib": m, "calibration": calib_info},