code/reasoning/ressoning_qwen3-30B-a3b_v2.py

import os
import json
import tqdm
import argparse
from openai import OpenAI
import re

# -----------------------------
#  CONFIGURATION
# -----------------------------
API_URL = "http://172.16.34.29:8004/v1" 
API_KEY = "EMPTY"
MODEL_NAME = "Qwen/Qwen3-30B-A3B-Instruct-2507" 

client = OpenAI(base_url=API_URL, api_key=API_KEY)

# -----------------------------
#  REASONING PROMPT
# -----------------------------
def reasoning_prompt(reference_text, statement, task_type="attribution"):
    if task_type == "attribution":
        # Checking if a summary subclaim is supported by the source medical text
        return f"""You are a senior clinical data validator. A previous system flagged a subclaim as 'not_supported' by the medical text. 
Verify if this is a False Negative.

### CONTEXT:
Medical Text (Source): {reference_text}
Subclaim (from Summary): {statement}

### TASK:
1. Search the Medical Text for paraphrased evidence or implicit support for the Subclaim.
2. Determine if it is 'supported' or 'not_supported'.

### OUTPUT FORMAT:
Provide internal reasoning in <think> tags, then conclude with exactly one word: 'supported' or 'not_supported'."""
    else:
        # Checking if a source fact is actually present in the summary (Completeness)
        return f"""You are a senior clinical data validator. A system flagged that a specific fact from the source medical text is missing ('not_supported') from the summary.
Verify if the summary actually contains this information.

### CONTEXT:
Summary Text: {reference_text}
Source Fact: {statement}

### TASK:
1. Search the Summary Text for the Source Fact. Look for synonyms or condensed mentions.
2. If the summary contains the info, label it 'supported'. If truly missing, label it 'not_supported'.

### OUTPUT FORMAT:
Provide internal reasoning in <think> tags, then conclude with exactly one word: 'supported' or 'not_supported'."""

# -----------------------------
#  LOGIC TO EXTRACT THINKING & LABEL
# -----------------------------
def get_reasoned_verdict(reference: str, statement: str, task_type: str):
    prompt = reasoning_prompt(reference, statement, task_type)
    
    try:
        response = client.chat.completions.create(
            model=MODEL_NAME,
            messages=[{"role": "user", "content": prompt}],
            temperature=0.1,
        )
        full_content = response.choices[0].message.content
        
        reasoning = ""
        if "<think>" in full_content and "</think>" in full_content:
            reasoning = re.search(r"<think>(.*?)</think>", full_content, re.DOTALL).group(1).strip()
            final_output = full_content.split("</think>")[-1].strip().lower()
        else:
            reasoning = "No explicit <think> tags provided."
            final_output = full_content.strip().lower()

        if "not_supported" in final_output:
            label = "not_supported"
        elif "supported" in final_output:
            label = "supported"
        else:
            label = "inconclusive"

        return reasoning, label

    except Exception as e:
        return str(e), "error_api"

# -----------------------------
#  MAIN PROCESSING
# -----------------------------
if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--input_file", type=str, required=True) 
    parser.add_argument("--save_path", type=str, default="/home/mshahidul/readctrl/data/reasoning/")
    args = parser.parse_args()

    os.makedirs(args.save_path, exist_ok=True)

    with open(args.input_file, "r") as f:
        data = json.load(f)
        
    save_filename = f"refined_v2_{os.path.basename(args.input_file)}"
    full_save_path = os.path.join(args.save_path, save_filename)
    
    print(f"Processing {len(data)} documents...")

    for doc in tqdm.tqdm(data):
        # We need the source text for Attribution and the summary text for Completeness
        # Assuming 'fulltext' is the source and 'summary' is the generated summary
        source_text = doc.get('fulltext', '')
        summary_text = doc.get('summary', '') # Ensure this key matches your JSON

        # 1. Audit Attribution Details
        if 'attribution_details' in doc:
            for item in doc['attribution_details']:
                if item.get('label') == "not_supported":
                    reasoning, new_label = get_reasoned_verdict(source_text, item.get('subclaim', ''), "attribution")
                    item['original_label'] = "not_supported"
                    item['reasoning_audit'] = reasoning
                    item['label'] = new_label
                    item['is_refined'] = True

        # 2. Audit Completeness Details
        if 'completeness_details' in doc:
            for item in doc['completeness_details']:
                if item.get('present_in_summary') == "not_supported":
                    # Here we check if the 'source_fact' is in the 'summary_text'
                    reasoning, new_label = get_reasoned_verdict(summary_text, item.get('source_fact', ''), "completeness")
                    item['original_label'] = "not_supported"
                    item['reasoning_audit'] = reasoning
                    item['present_in_summary'] = new_label
                    item['is_refined'] = True

        # Save state periodically
        with open(full_save_path, "w") as f:
            json.dump(data, f, indent=2, ensure_ascii=False)

    print(f"Refinement complete. Saved to {full_save_path}")