code/classifier/classifier.py

import os
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = "2"

import torch
from unsloth import FastLanguageModel
import json
import tqdm
import re

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

def load_finetuned_model(model_path: str):
    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, # Set to True if you want 4bit inference for speed/memory
        load_in_8bit=False,
        full_finetuning=False,
    )
    # Enable native 2x faster inference
    FastLanguageModel.for_inference(model)
    
    _model_cache["model"], _model_cache["tokenizer"] = model, tokenizer
    return model, tokenizer

# -----------------------------
#  READABILITY CLASSIFICATION PROMPT
# -----------------------------
def classification_prompt(full_text: str, summary: str) -> str:
    """
    Constructs the prompt to classify readability of the summary
    based on the context of the full text.
    """
    prompt = f"""You are a medical readability evaluator.

### Task
Compare the "GENERATED TEXT" against the "FULL TEXT" to determine its readability for a general, non-medical audience.

### Input Data
- **FULL TEXT:** {full_text}
- **GENERATED TEXT (Evaluate this):** {summary}

### Readability Scale
1: Very Easy - Minimal medical language, uses simple terms.
2: Easy - Accessible to most, minor jargon explained.
3: Medium - Some technical terms, moderate complexity.
4: Hard - Clinical tone, assumes some prior knowledge.
5: Very Hard - Extremely technical, requires medical expertise.

### Constraints
- Evaluate ONLY the "GENERATED TEXT".
- Use "FULL TEXT" only for context of the subject matter.
- Do NOT assess factual accuracy.

### Output Format
Return ONLY a valid JSON object:
{{
  "readability_score": <integer_1_to_5>
}}"""
    return prompt

# -----------------------------
#  INFERENCE FUNCTION
# -----------------------------
def infer_readability(full_text: str, 
                      summary: str,
                      model_path: str) -> dict:
    
    model, tokenizer = load_finetuned_model(model_path)
    prompt = classification_prompt(full_text, summary)

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

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

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

    with torch.no_grad():
        output_ids = model.generate(
            **inputs,
            max_new_tokens=50, # Classification only needs a few tokens
            temperature=0.1,   # Low temperature for classification consistency
            do_sample=False,
        )

    output_text = tokenizer.decode(output_ids[0][len(inputs.input_ids[0]):], skip_special_tokens=True).strip()

    # Clean up output (remove thinking or markdown)
    if "</think>" in output_text:
        output_text = output_text.split("</think>")[-1].strip()
    
    # Simple regex to extract JSON if the model adds conversational filler
    try:
        match = re.search(r"\{.*\}", output_text, re.DOTALL)
        if match:
            return json.loads(match.group())
        return {"readability_score": "error", "raw": output_text}
    except Exception:
        return {"readability_score": "error", "raw": output_text}

# -----------------------------
#  MAIN EXECUTION
# -----------------------------
if __name__ == "__main__":
    # Settings based on your paths
    INPUT_FILE = "/home/mshahidul/readctrl/data/processed_raw_data/multiclinsum_test_en.json"
    SAVE_FOLDER = "/home/mshahidul/readctrl/data/classified_readability"
    # Note: Ensure this path points to your CLASSIFIER model, not the subclaim extractor
    MODEL_PATH = "/home/mshahidul/readctrl_model/qwen3-32B_classifier_en" 

    os.makedirs(SAVE_FOLDER, exist_ok=True)
    file_name = os.path.basename(INPUT_FILE).split(".json")[0]
    OUTPUT_FILE = os.path.join(SAVE_FOLDER, f"classified_{file_name}.json")

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

    # 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}

    print(f"Starting classification. Saving to: {OUTPUT_FILE}")

    for item in tqdm.tqdm(data):
        if item["id"] in existing_ids:
            continue

        full_text = item.get("fulltext", "")
        summary = item.get("summary", "")

        classification_res = infer_readability(
            full_text=full_text,
            summary=summary,
            model_path=MODEL_PATH
        )

        result.append({
            "id": item["id"],
            "readability_score": classification_res.get("readability_score"),
            "fulltext": full_text,
            "summary": summary
        })

        # Checkpoint every 50 items
        if len(result) % 50 == 0:
            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(f"Classification completed. {len(result)} items processed.")