File size: 7,015 Bytes
9df97a2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#!/usr/bin/env python3
"""Fine-tune a recruiter-domain NER model from annotated CV data.

Expected input formats:
- JSONL with tokens + ner_tags
- JSON with records containing `tokens` and `ner_tags`

This script is intentionally defensive: if `transformers` is unavailable,
it still validates and exports a prepared dataset snapshot for later training.
"""

from __future__ import annotations

import argparse
import json
from pathlib import Path
from typing import Any, Dict, List, Sequence, Tuple

try:
    from datasets import Dataset  # type: ignore
    from transformers import AutoModelForTokenClassification, AutoTokenizer, Trainer, TrainingArguments  # type: ignore

    TRANSFORMERS_AVAILABLE = True
except Exception:
    Dataset = None
    AutoModelForTokenClassification = None
    AutoTokenizer = None
    Trainer = None
    TrainingArguments = None
    TRANSFORMERS_AVAILABLE = False

LABEL_NAMES = ["O", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-SKILL", "I-SKILL", "B-EDU", "I-EDU"]
DEFAULT_MODEL = "dslim/bert-base-NER"


def load_annotations(path: Path) -> List[Dict[str, Any]]:
    """Load token/label annotations from JSON or JSONL."""
    if not path.exists():
        raise FileNotFoundError(f"Annotation file not found: {path}")

    text = path.read_text(encoding="utf-8").strip()
    if not text:
        return []

    if path.suffix.lower() == ".jsonl":
        records = []
        for line in text.splitlines():
            if line.strip():
                records.append(json.loads(line))
        return records

    payload = json.loads(text)
    if isinstance(payload, list):
        return payload
    if isinstance(payload, dict) and isinstance(payload.get("data"), list):
        return payload["data"]
    raise ValueError("Unsupported annotation format")


def validate_records(records: Sequence[Dict[str, Any]]) -> List[Dict[str, Any]]:
    """Keep only records with aligned tokens/labels."""
    valid_records: List[Dict[str, Any]] = []
    for record in records:
        tokens = record.get("tokens")
        ner_tags = record.get("ner_tags")
        if not isinstance(tokens, list) or not isinstance(ner_tags, list):
            continue
        if len(tokens) != len(ner_tags) or not tokens:
            continue
        valid_records.append({"tokens": [str(token) for token in tokens], "ner_tags": [str(tag) for tag in ner_tags]})
    return valid_records


def build_label_map(records: Sequence[Dict[str, Any]]) -> Tuple[Dict[str, int], Dict[int, str]]:
    labels = {"O"}
    for record in records:
        labels.update(str(tag) for tag in record.get("ner_tags", []))
    ordered_labels = ["O"] + sorted(label for label in labels if label != "O")
    label2id = {label: index for index, label in enumerate(ordered_labels)}
    id2label = {index: label for label, index in label2id.items()}
    return label2id, id2label


def tokenize_and_align_labels(examples: Dict[str, List[Any]], tokenizer, label2id: Dict[str, int]) -> Dict[str, Any]:
    tokenized_inputs = tokenizer(
        examples["tokens"],
        is_split_into_words=True,
        truncation=True,
        padding=False,
    )

    aligned_labels = []
    for batch_index, labels in enumerate(examples["ner_tags"]):
        word_ids = tokenized_inputs.word_ids(batch_index=batch_index)
        previous_word_id = None
        label_ids: List[int] = []
        for word_id in word_ids:
            if word_id is None:
                label_ids.append(-100)
            elif word_id != previous_word_id:
                label_ids.append(label2id.get(labels[word_id], label2id["O"]))
            else:
                label = labels[word_id]
                if label.startswith("B-"):
                    label = "I-" + label[2:]
                label_ids.append(label2id.get(label, label2id["O"]))
            previous_word_id = word_id
        aligned_labels.append(label_ids)

    tokenized_inputs["labels"] = aligned_labels
    return tokenized_inputs


def export_prepared_dataset(records: Sequence[Dict[str, Any]], output_dir: Path) -> Path:
    output_dir.mkdir(parents=True, exist_ok=True)
    export_path = output_dir / "recruiter_ner_prepared_dataset.json"
    export_path.write_text(json.dumps(list(records), indent=2, ensure_ascii=False), encoding="utf-8")
    return export_path


def train_model(records: Sequence[Dict[str, Any]], model_name: str, output_dir: Path, epochs: int) -> Path:
    if not TRANSFORMERS_AVAILABLE:
        raise RuntimeError("transformers/datasets are not available in this environment")

    label2id, id2label = build_label_map(records)
    tokenizer = AutoTokenizer.from_pretrained(model_name)
    model = AutoModelForTokenClassification.from_pretrained(
        model_name,
        num_labels=len(label2id),
        label2id=label2id,
        id2label=id2label,
    )

    dataset = Dataset.from_list(list(records))
    tokenized_dataset = dataset.map(lambda batch: tokenize_and_align_labels(batch, tokenizer, label2id), batched=True)

    output_dir.mkdir(parents=True, exist_ok=True)
    training_args = TrainingArguments(
        output_dir=str(output_dir),
        learning_rate=2e-5,
        per_device_train_batch_size=8,
        per_device_eval_batch_size=8,
        num_train_epochs=epochs,
        weight_decay=0.01,
        logging_steps=10,
        save_strategy="epoch",
        report_to=[],
    )

    trainer = Trainer(
        model=model,
        args=training_args,
        train_dataset=tokenized_dataset,
        tokenizer=tokenizer,
    )
    trainer.train()
    trainer.save_model(str(output_dir))
    tokenizer.save_pretrained(str(output_dir))
    return output_dir


def main() -> int:
    parser = argparse.ArgumentParser(description="Fine-tune recruiter NER model")
    parser.add_argument("--annotations", required=True, help="Path to JSON/JSONL annotation file")
    parser.add_argument("--output-dir", default="models/recruiter_ner", help="Training output directory")
    parser.add_argument("--model-name", default=DEFAULT_MODEL, help="Base HF model to fine-tune")
    parser.add_argument("--epochs", type=int, default=3, help="Training epochs")
    parser.add_argument("--export-only", action="store_true", help="Prepare dataset without training")
    args = parser.parse_args()

    annotation_path = Path(args.annotations).expanduser().resolve()
    output_dir = Path(args.output_dir).expanduser().resolve()

    records = validate_records(load_annotations(annotation_path))
    if not records:
        raise SystemExit("No valid annotation records found")

    export_path = export_prepared_dataset(records, output_dir)
    print(f"Prepared dataset exported to {export_path}")

    if args.export_only or not TRANSFORMERS_AVAILABLE:
        print("Training skipped: exporting only or transformers unavailable")
        return 0

    trained_dir = train_model(records, args.model_name, output_dir, args.epochs)
    print(f"Model trained and saved to {trained_dir}")
    return 0


if __name__ == "__main__":
    raise SystemExit(main())