diff --git a/train_all.py b/train_all.py index 3d45c15..56272f0 100644 --- a/train_all.py +++ b/train_all.py @@ -50,20 +50,30 @@ def bisect_auc(probs, labels): 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): +def augment_crosspair_negs(rows, win_tokens, max_new=3200): """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。""" + 『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。 + canary_en 诊断: 同域 hard negative(同产品不同主题)是英文难点 → en 候选优先填满 50%。""" 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) + cand_en = [i for i in cand if _lang(rows[i].get("doc_text") or "") == "en"] + cand_zh = [i for i in cand if _lang(rows[i].get("doc_text") or "") == "zh"] + # en 优先: en 填一半, 剩余 zh (en 不足时 zh 补齐) + n_en = min(len(cand_en), max_new // 2) + n_zh = min(len(cand_zh), max_new - n_en) + picked = cand_en[:n_en] + cand_zh[:n_zh] + if n_en + n_zh < max_new: + rest = cand_en[n_en:] + cand_zh[n_zh:] + picked += rest[:max_new - n_en - n_zh] pools = {} for i, r in enumerate(rows): pools.setdefault(_lang(r["query"]), []).append(i) added = [] - for i in cand[:max_new]: + for i in picked: L = _lang(rows[i].get("doc_text") or "") pool = pools[L] j = rng.choice(pool) @@ -75,6 +85,19 @@ def augment_crosspair_negs(rows, win_tokens, max_new=2500): "hard_label": 0, "soft_label": 0.0}) return added +def oversample_en_pos(rows, factor=1.5): + """canary_en 弱判别: en 正例语义学习欠充分 → 过采样 ×factor (粗放但便宜)""" + en_pos = [r for r in rows + if r.get("hard_label") == 1 and _lang(r["query"]) == "en"] + extra_n = int(len(en_pos) * (factor - 1)) + rng = random.Random(42) + extra = [] + for r in rng.sample(en_pos, min(extra_n, len(en_pos))): + d = dict(r) + d["pair_id"] = f"os-{d.get('pair_id', '?')}" + extra.append(d) + return extra + def phase_data(): """v05n tar.gz 分片重组 → DATA_DIR/train.jsonl""" import glob, shutil, tarfile @@ -237,20 +260,22 @@ def run(args): self.q = {"t": "noul", "ins": INSTR, "crit": None} self.rng = random.Random(20261105) def _doc_window(self, text): + """返回 (window_text, aug_flag)。aug_p 概率对长文档随机取窗(答案可能不在窗内)""" if not self.train_aug: - return text + return text, False ids = tok(text[:12000], add_special_tokens=False)["input_ids"] if len(ids) <= self.win_tokens: - return text + return text, False 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 + return tok.decode(ids[s:s + self.win_tokens]), True + return text, False 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{self._doc_window(r['doc_text'])}" + doc_win, aug = self._doc_window(r["doc_text"]) + state = f"用户问题: {r['query']}\n\n文档内容:\n{doc_win}" 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)) @@ -258,12 +283,18 @@ def run(args): if soft is None: soft = r["hard_label"] soft_clean = float(soft) if soft in (0, 1) else -1.0 + # 窗口标签平滑: 随机窗口的正例, 答案可能不在窗内 → BCE 目标 0.85 + label = float(r["hard_label"]) + if aug and label == 1.0: + label = 0.85 + if soft_clean < 0: + soft_clean = 0.85 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"])), + "label": torch.tensor(label), "soft": torch.tensor(float(soft)), "soft_clean": torch.tensor(soft_clean)} @@ -303,9 +334,12 @@ def run(args): 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) + xp = augment_crosspair_negs(train, args.win_tokens, max_new=3200) train = train + xp - log(f"crosspair long negs: +{len(xp)} (total {len(train)})") + os_rows = oversample_en_pos(train, factor=args.en_boost) + train = train + os_rows + log(f"crosspair long negs: +{len(xp)}, en_pos oversample: +{len(os_rows)} " + f"(total {len(train)})") log(f"train={len(train)} calib={len(calib)}") os.makedirs(OUT_DIR, exist_ok=True) @@ -569,4 +603,5 @@ if __name__ == "__main__": 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") + ap.add_argument("--en_boost", type=float, default=1.5, help="en 正例过采样倍数 (1=关)") run(ap.parse_args())