File size: 11,868 Bytes
f29b6e6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
#!/usr/bin/env python3
# train.py
#
# Fine-tune mdeberta-v3-base for binary token classification (LINK vs O).
# Inputs (same dir): train_windows.jsonl, val_windows.jsonl  (from prep.py)
# JSONL per line: {"doc_id": int, "window_id": int, "input_ids": [...], "attention_mask": [...], "labels": [...]}

import os
import json
import argparse
import inspect
from typing import List, Dict, Any

import numpy as np
import torch
from torch.utils.data import Dataset

from transformers import (
    AutoTokenizer,
    AutoConfig,
    AutoModelForTokenClassification,
    Trainer,
    TrainingArguments,
    set_seed,
)

# --------------------------
# Dataset
# --------------------------

class JsonlTokenDataset(Dataset):
    """Loads JSONL produced by prep.py. Masks special tokens in labels to -100."""
    def __init__(self, path: str, tokenizer: AutoTokenizer):
        self.path = path
        self.tokenizer = tokenizer
        self.samples: List[Dict[str, Any]] = []
        with open(self.path, "r", encoding="utf-8") as f:
            for line in f:
                rec = json.loads(line)
                self.samples.append(rec)

        # Mask specials to -100
        for rec in self.samples:
            input_ids = rec["input_ids"]
            labels = rec["labels"]
            try:
                special_mask = tokenizer.get_special_tokens_mask(input_ids, already_has_special_tokens=True)
            except Exception:
                spec = set(tokenizer.all_special_ids or [])
                special_mask = [1 if t in spec else 0 for t in input_ids]
            rec["labels"] = [-100 if sm == 1 else int(l) for l, sm in zip(labels, special_mask)]

    def __len__(self): return len(self.samples)

    def __getitem__(self, idx: int) -> Dict[str, torch.Tensor]:
        r = self.samples[idx]
        return {
            "input_ids": torch.tensor(r["input_ids"], dtype=torch.long),
            "attention_mask": torch.tensor(r["attention_mask"], dtype=torch.long),
            "labels": torch.tensor(r["labels"], dtype=torch.long),
        }

# --------------------------
# Collator (pads inputs + labels)
# --------------------------

class SimpleTokenCollator:
    """Pads input_ids with pad_token_id, attention_mask with 0, labels with -100."""
    def __init__(self, tokenizer: AutoTokenizer, pad_to_multiple_of: int = None):
        self.pad_id = tokenizer.pad_token_id if tokenizer.pad_token_id is not None else 0
        self.pad_to_multiple = pad_to_multiple_of

    def __call__(self, features: List[Dict[str, torch.Tensor]]) -> Dict[str, torch.Tensor]:
        ids = [f["input_ids"].tolist() for f in features]
        att = [f["attention_mask"].tolist() for f in features]
        lab = [f["labels"].tolist() for f in features]
        max_len = max(len(x) for x in ids)
        if self.pad_to_multiple and max_len % self.pad_to_multiple != 0:
            max_len = ((max_len // self.pad_to_multiple) + 1) * self.pad_to_multiple

        def pad(seq, val): return seq + [val] * (max_len - len(seq))
        ids = [pad(x, self.pad_id) for x in ids]
        att = [pad(x, 0) for x in att]
        lab = [pad(x, -100) for x in lab]
        return {
            "input_ids": torch.tensor(ids, dtype=torch.long),
            "attention_mask": torch.tensor(att, dtype=torch.long),
            "labels": torch.tensor(lab, dtype=torch.long),
        }

# --------------------------
# Class weights
# --------------------------

def compute_class_weights(dataset: JsonlTokenDataset) -> torch.Tensor:
    pos = 0; neg = 0
    for rec in dataset.samples:
        for l in rec["labels"]:
            if l == -100: continue
            if l == 1: pos += 1
            else: neg += 1
    return torch.tensor([1.0, (neg / max(1, pos)) if pos > 0 else 1.0], dtype=torch.float)

# --------------------------
# Metrics
# --------------------------

def compute_metrics_fn(eval_pred):
    logits, labels = eval_pred
    preds = np.argmax(logits, axis=-1)
    y_true, y_pred = [], []
    for p, l in zip(preds, labels):
        for pi, li in zip(p, l):
            if li == -100: continue
            y_true.append(int(li)); y_pred.append(int(pi))
    if not y_true:
        return {"accuracy":0.0,"precision":0.0,"recall":0.0,"f1":0.0,"pos_rate_true":0.0,"pos_rate_pred":0.0}
    y_true = np.array(y_true); y_pred = np.array(y_pred)
    tp = int(np.sum((y_pred==1)&(y_true==1))); fp = int(np.sum((y_pred==1)&(y_true==0)))
    tn = int(np.sum((y_pred==0)&(y_true==0))); fn = int(np.sum((y_pred==0)&(y_true==1)))
    acc = (tp+tn)/max(1,tp+tn+fp+fn); prec = tp/max(1,tp+fp); rec = tp/max(1,tp+fn)
    f1 = (2*prec*rec/max(1e-12,prec+rec)) if (prec+rec)>0 else 0.0
    return {"accuracy":acc,"precision":prec,"recall":rec,"f1":f1,
            "pos_rate_true":float(np.mean(y_true)),"pos_rate_pred":float(np.mean(y_pred))}

# --------------------------
# Weighted Trainer (accepts extra kwargs like num_items_in_batch)
# --------------------------

class WeightedCELossTrainer(Trainer):
    def __init__(self, class_weights: torch.Tensor = None, **kwargs):
        super().__init__(**kwargs); self.class_weights = class_weights

    def compute_loss(self, model, inputs, return_outputs=False, **kwargs):
        labels = inputs["labels"]
        outputs = model(input_ids=inputs["input_ids"], attention_mask=inputs["attention_mask"])
        logits = outputs.logits  # [B,T,2]
        loss_fct = torch.nn.CrossEntropyLoss(
            weight=(self.class_weights.to(logits.device) if self.class_weights is not None else None)
        )
        mask = labels.ne(-100)
        loss = loss_fct(logits.view(-1,2)[mask.view(-1)], labels.view(-1)[mask.view(-1)])
        return (loss, outputs) if return_outputs else loss

# --------------------------
# TrainingArguments compatibility
# --------------------------

def build_training_arguments(args) -> TrainingArguments:
    sig = set(inspect.signature(TrainingArguments.__init__).parameters.keys())

    supports_eval_strategy  = "evaluation_strategy" in sig
    supports_save_strategy  = "save_strategy" in sig
    supports_log_strategy   = "logging_strategy" in sig
    supports_report_to      = "report_to" in sig
    supports_load_best      = "load_best_model_at_end" in sig
    supports_metric_forbest = "metric_for_best_model" in sig
    supports_workers        = "dataloader_num_workers" in sig

    kw = {
        "output_dir": args.output_dir,
        "num_train_epochs": args.epochs,
        "per_device_train_batch_size": args.train_batch_size,
        "per_device_eval_batch_size": args.eval_batch_size,
        "learning_rate": args.lr,
        "weight_decay": args.weight_decay,
        "logging_steps": args.logging_steps,
        "eval_steps": args.eval_steps,
        "save_steps": args.save_steps,
        "save_total_limit": 2,
        "seed": args.seed,
        "gradient_accumulation_steps": args.gradient_accumulation_steps,
        "fp16": args.fp16,
        "bf16": args.bf16,
        "gradient_checkpointing": args.gradient_checkpointing,
        "log_level": "info",
    }
    if supports_workers:
        kw["dataloader_num_workers"] = args.num_workers
    if supports_report_to:
        kw["report_to"] = (None if args.report_to == "none" else ["wandb"])

    # Pair strategies only if BOTH are supported to avoid mismatches
    if supports_eval_strategy and supports_save_strategy:
        kw["evaluation_strategy"] = "steps"
        kw["save_strategy"] = "steps"
        if supports_log_strategy:
            kw["logging_strategy"] = "steps"
        if supports_load_best:
            kw["load_best_model_at_end"] = True
        if supports_metric_forbest:
            kw["metric_for_best_model"] = "f1"
            if "greater_is_better" in sig:
                kw["greater_is_better"] = True
    else:
        for k in ("evaluation_strategy","save_strategy","logging_strategy","load_best_model_at_end",
                  "metric_for_best_model","greater_is_better"):
            kw.pop(k, None)
        if "evaluate_during_training" in sig and args.eval_steps > 0:
            kw["evaluate_during_training"] = True

    kw = {k: v for k, v in kw.items() if k in sig}
    return TrainingArguments(**kw)

# --------------------------
# Args
# --------------------------

def parse_args():
    ap = argparse.ArgumentParser(description="Train binary token classification model for link anchors.")
    ap.add_argument("--model_name", default="microsoft/mdeberta-v3-base", help="HF model name or local path.")
    ap.add_argument("--train_path", default="train_windows.jsonl", help="Training JSONL.")
    ap.add_argument("--val_path", default="val_windows.jsonl", help="Validation JSONL.")
    ap.add_argument("--output_dir", default="model_link_token_cls", help="Output directory.")

    ap.add_argument("--epochs", type=int, default=3)
    ap.add_argument("--lr", type=float, default=2e-5)
    ap.add_argument("--weight_decay", type=float, default=0.01)
    ap.add_argument("--train_batch_size", type=int, default=16)
    ap.add_argument("--eval_batch_size", type=int, default=32)
    ap.add_argument("--logging_steps", type=int, default=50)
    ap.add_argument("--eval_steps", type=int, default=500)
    ap.add_argument("--save_steps", type=int, default=500)

    ap.add_argument("--seed", type=int, default=42)
    ap.add_argument("--gradient_accumulation_steps", type=int, default=1)
    ap.add_argument("--fp16", action="store_true")
    ap.add_argument("--bf16", action="store_true")
    ap.add_argument("--gradient_checkpointing", action="store_true")
    ap.add_argument("--report_to", default="wandb", choices=["wandb","none"])
    ap.add_argument("--pad_to_multiple_of", type=int, default=8)
    ap.add_argument("--num_workers", type=int, default=2)
    return ap.parse_args()

# --------------------------
# Main
# --------------------------

def main():
    args = parse_args()
    set_seed(args.seed)

    tokenizer = AutoTokenizer.from_pretrained(args.model_name, use_fast=True)

    train_ds = JsonlTokenDataset(args.train_path, tokenizer)
    val_ds   = JsonlTokenDataset(args.val_path, tokenizer)

    id2label = {0: "O", 1: "LINK"}
    label2id = {"O": 0, "LINK": 1}
    config = AutoConfig.from_pretrained(args.model_name, num_labels=2, id2label=id2label, label2id=label2id)
    model  = AutoModelForTokenClassification.from_pretrained(args.model_name, config=config)

    class_weights = compute_class_weights(train_ds)

    collator = SimpleTokenCollator(
        tokenizer=tokenizer,
        pad_to_multiple_of=(args.pad_to_multiple_of if torch.cuda.is_available() else None),
    )

    training_args = build_training_arguments(args)

    trainer = WeightedCELossTrainer(
        model=model,
        args=training_args,
        train_dataset=train_ds,
        eval_dataset=val_ds,
        data_collator=collator,
        tokenizer=tokenizer,
        compute_metrics=compute_metrics_fn,
        class_weights=class_weights,
    )

    trainer.train()
    trainer.save_model(args.output_dir)
    tokenizer.save_pretrained(args.output_dir)

    metrics = trainer.evaluate()
    trainer.log_metrics("eval", metrics)
    trainer.save_metrics("eval", metrics)
    trainer.save_state()

    with open(os.path.join(args.output_dir, "label_map.json"), "w", encoding="utf-8") as f:
        json.dump({"0":"O","1":"LINK"}, f)

    print("=== Training complete ===")
    print(f"Output dir: {args.output_dir}")
    print(f"Class weights [neg, pos]: [{class_weights[0].item():.4f}, {class_weights[1].item():.4f}]")
    print(f"Eval metrics: {metrics}")

if __name__ == "__main__":
    main()