code/RL_model/verl/Search-R1/dataset/data_prep.py

import os
import json
import datasets
import argparse
from verl.utils.hdfs_io import copy, makedirs

# 1. Define the exact Prompt Template from your requirements
# /home/mshahidul/readctrl/code/RL_model/verl/verl_train/dataset/prompt
with open("/home/mshahidul/readctrl/code/RL_model/verl/verl_train/dataset/prompt", 'r') as f:
    PROMPT_TEMPLATE = f.read()

def make_map_fn(split, data_source):
    def process_fn(example, idx):
        # Extract fields from your specific JSON keys: ['id', 'fulltext', 'summary']
        full_text = example.pop('fulltext')
        gold_summary = example.pop('summary')
        
        # Format the prompt using your template
        # Note: Added 'English' as default source lang based on filename
        prompt_content = PROMPT_TEMPLATE.format(
            source_lang="English",
            gold_summary=gold_summary,
            full_text=full_text
        )

        return {
            "data_source": data_source,
            "prompt": [{
                "role": "user",
                "content": prompt_content
            }],
            "ability": "summarization",
            "reward_model": {
                "style": "rule",
                "ground_truth": gold_summary 
            },
            "extra_info": {
                "split": split,
                "index": idx,
                "original_id": example.get('id', idx)
            }
        }
    return process_fn

if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    # Path to your input JSON
    parser.add_argument('--input_path', default='/home/mshahidul/readctrl/data/processed_test_raw_data/multiclinsum_test_en.json')
    # Updated destination as requested
    parser.add_argument('--local_dir', default='/home/mshahidul/readctrl/code/RL_model/verl/Search-R1/dataset')
    args = parser.parse_args()

    data_source = 'multiclinsum'

    # Load your local JSON file
    with open(args.input_path, 'r') as f:
        raw_data = json.load(f)
    
    # Convert to HuggingFace Dataset
    dataset = datasets.Dataset.from_list(raw_data)

    # Split into train/test (95% train, 5% test)
    split_dataset = dataset.train_test_split(test_size=0.05, seed=42)

    # Apply the mapping transformation for each split
    processed_train = split_dataset["train"].map(
        function=make_map_fn('train', data_source),
        with_indices=True
    )
    processed_test = split_dataset["test"].map(
        function=make_map_fn('test', data_source),
        with_indices=True
    )

    # Create the directory if it doesn't exist
    os.makedirs(args.local_dir, exist_ok=True)

    # Save to Parquet in the specified location
    train_output_path = os.path.join(args.local_dir, 'train.parquet')
    test_output_path = os.path.join(args.local_dir, 'test.parquet')
    processed_train.to_parquet(train_output_path)
    processed_test.to_parquet(test_output_path)

    print(f"--- Dataset Preparation Complete ---")
    print(f"Train file saved to: {train_output_path}")
    print(f"Test file saved to: {test_output_path}")
    print(f"Total train records: {len(processed_train)}")
    print(f"Total test records: {len(processed_test)}")