File size: 4,407 Bytes
9c6961c | 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 | import os
import json
import tqdm
import argparse
from openai import OpenAI
# -----------------------------
# CONFIGURATION
# -----------------------------
# Ensure this matches the model path used in your run_vllm.sh script
MODEL_NAME = "/home/mshahidul/readctrl_model/full_model/qwen3-32B_subclaims_BF16_merged"
API_URL = "http://localhost:8015/v1"
API_KEY = "EMPTY" # vLLM requires a key, but it can be anything if not set on server
# Initialize Client
client = OpenAI(base_url=API_URL, api_key=API_KEY)
# -----------------------------
# 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 (vLLM)
# -----------------------------
def infer_subclaims(medical_text: str, temperature: float = 0.2) -> str:
"""Sends prompt to vLLM server and returns generated text."""
# 1. Prepare the prompt
final_prompt = extraction_prompt(medical_text)
# 2. Call the vLLM Server via OpenAI API
try:
response = client.chat.completions.create(
model=MODEL_NAME,
messages=[
{"role": "user", "content": final_prompt}
],
max_tokens=1000, # Limit generation length
temperature=temperature,
top_p=0.9,
frequency_penalty=0.0,
presence_penalty=0.0,
)
res = response.choices[0].message.content.strip()
res=res.split("</think>")[-1].strip()
return res
except Exception as e:
print(f"Error during API call: {e}")
return None
# -----------------------------
# MAIN EXECUTION
# -----------------------------
if __name__ == "__main__":
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"
os.makedirs(SAVE_FOLDER, exist_ok=True)
OUTPUT_FILE = os.path.join(SAVE_FOLDER, f"extracted_subclaims_{file_name}.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:
try:
result = json.load(f)
except json.JSONDecodeError:
result = []
existing_ids = {item["id"] for item in result}
print(f"Starting inference on {len(data)} items using vLLM server...")
save=False
# --------------------------------------------------------
# PROCESS EACH MEDICAL TEXT
# --------------------------------------------------------
for item in tqdm.tqdm(data):
if item["id"] in existing_ids:
continue
medical_text = item.get("fulltext", "")
# Call the vLLM inference function
extracted = infer_subclaims(medical_text)
result.append({
"id": item["id"],
"medical_text": medical_text,
"subclaims": extracted,
"summary": item.get("summary", "")
})
# Save every 20 entries
if len(result) % 20 == 0:
with open(OUTPUT_FILE, "w") as f:
if save:
json.dump(result, f, indent=4, ensure_ascii=False)
# Final save
with open(OUTPUT_FILE, "w") as f:
if save:
json.dump(result, f, indent=4, ensure_ascii=False)
print(f"Extraction completed. Saved to {OUTPUT_FILE}") |