File size: 4,822 Bytes
030876e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import os
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = "1"

import torch
from unsloth import FastLanguageModel
import json
import tqdm

# -----------------------------
#  MODEL CACHE
# -----------------------------
_model_cache = {"model": None, "tokenizer": None}

def load_finetuned_model(model_path: str):
    """Load and cache your fine-tuned subclaim extraction model + tokenizer."""
    if _model_cache["model"] is not None:
        return _model_cache["model"], _model_cache["tokenizer"]

    model, tokenizer = FastLanguageModel.from_pretrained(
        model_name=model_path,
        max_seq_length=8192,
        load_in_4bit=False,
        load_in_8bit=False,
        full_finetuning=False,
    )
    _model_cache["model"], _model_cache["tokenizer"] = model, tokenizer
    return model, tokenizer


# -----------------------------
#  SUBCLAIM EXTRACTION PROMPT
# -----------------------------
def extraction_prompt(medical_text: str) -> str:
    prompt = f"""
You are an expert medical annotator. Your task is to extract granular, factual subclaims from medical text.
A subclaim is the smallest standalone factual unit that can be independently verified.

Instructions:
1. Read the provided medical text.
2. Break it into clear, objective, atomic subclaims.
3. Each subclaim must come directly from the text.
4. Do not add, guess, or infer information.
5. Each subclaim should be short, specific, and verifiable.
6. Return ONLY a Python-style list of strings.

Medical Text:
{medical_text}

Return your output in JSON list format, like:
[
  "subclaim 1",
  "subclaim 2",
  ...
]
"""
    return prompt


# -----------------------------
#  INFERENCE FUNCTION
# -----------------------------
def infer_subclaims(medical_text: str,
                    model_path: str,
                    temperature: float = 0.2) -> str:
    
    model, tokenizer = load_finetuned_model(model_path)

    prompt = extraction_prompt(medical_text)

    messages = [{"role": "user", "content": prompt}]

    chat_text = tokenizer.apply_chat_template(
        messages,
        tokenize=False,
        add_generation_prompt=True,
        enable_thinking=False,
    )

    inputs = tokenizer(chat_text, return_tensors="pt").to("cuda")

    with torch.no_grad():
        output_ids = model.generate(
            **inputs,
            max_new_tokens=512,
            temperature=temperature,
            top_p=0.9,
            top_k=10,
            do_sample=False,
        )

    output_text = tokenizer.decode(output_ids[0], skip_special_tokens=True).strip()

    # Remove thinking if model inserts `<think>`
    if "</think>" in output_text:
        output_text = output_text.split("</think>")[-1].strip()

    return output_text


# -----------------------------
#  MAIN EXECUTION
# -----------------------------
if __name__ == "__main__":
    import argparse
    parser = argparse.ArgumentParser()
    parser.add_argument("--input_file", type=str, required=True,
                        help="Path to the input JSON file containing medical texts.")
    args = parser.parse_args()
    INPUT_FILE = args.input_file
    file_name=os.path.basename(INPUT_FILE).split(".json")[0]
    SAVE_FOLDER = "/home/mshahidul/readctrl/data/extracting_subclaim"
    MODEL_PATH = "/home/mshahidul/readctrl_model/qwen3-32B_subclaims-extraction-8b_ctx"

    os.makedirs(SAVE_FOLDER, exist_ok=True)

    OUTPUT_FILE = os.path.join(SAVE_FOLDER, f"extracted_subclaims_{file_name}_en.json")

    # Load input dataset
    with open(INPUT_FILE, "r") as f:
        data = json.load(f)

    # Load existing results (resume mode)
    result = []
    if os.path.exists(OUTPUT_FILE):
        with open(OUTPUT_FILE, "r") as f:
            result = json.load(f)

    existing_ids = {item["id"] for item in result}

    # --------------------------------------------------------
    # PROCESS EACH MEDICAL TEXT
    # --------------------------------------------------------
    for item in tqdm.tqdm(data):
        if item["id"] in existing_ids:
            continue

        medical_text = item.get("fulltext", "")

        extracted = infer_subclaims(
            medical_text,
            model_path=MODEL_PATH
        )

        result.append({
            "id": item["id"],
            "medical_text": medical_text,
            "subclaims": extracted,
            "summary": item.get("summary", "")
        })

        # Save every 20 entries
        if len(result) % 20 == 0:
            print(f"Saving intermediate results... Total processed: {len(result)}")
            with open(OUTPUT_FILE, "w") as f:
                json.dump(result, f, indent=4, ensure_ascii=False)

    # Final save
    with open(OUTPUT_FILE, "w") as f:
        json.dump(result, f, indent=4, ensure_ascii=False)

    print("Extraction completed.")