src/prepare_reasoning_dataset.py

import json
import os
from datasets import load_dataset
import sys
sys.path.append(os.path.dirname(os.path.abspath(__file__)))
import config

def format_conversations(conversations):
    text_parts = []
    system_prompt = "You are an advanced AI reasoning assistant powered by a Selective State Space Model. You solve complex problems step-by-step using internal reasoning traces wrapped in <think>...</think> tags."
    text_parts.append(f"<|im_start|>system\n{system_prompt}<|im_end|>\n")
    
    for msg in conversations:
        role = msg.get("from", "user")
        if role in ["human", "user"]:
            role_name = "user"
        else:
            role_name = "assistant"
        content = msg.get("value", "")
        text_parts.append(f"<|im_start|>{role_name}\n{content}<|im_end|>\n")
        
    return "".join(text_parts)

def main():
    print(f"Loading reasoning dataset: {config.DATASET_ID}...")
    ds = load_dataset(config.DATASET_ID, split="train")
    
    os.makedirs(config.DATA_DIR, exist_ok=True)
    os.makedirs(config.RESULTS_DIR, exist_ok=True)
    os.makedirs(config.DOCS_DIR, exist_ok=True)
    
    total = len(ds)
    print(f"Loaded {total} samples. Processing and split into train, val, eval_100...")
    
    processed = []
    eval_prompts = []
    
    for idx, sample in enumerate(ds):
        convs = sample.get("conversations", [])
        if not convs:
            continue
            
        full_text = format_conversations(convs)
        processed.append({"text": full_text})
        
        if len(eval_prompts) < 100:
            user_msg = ""
            for msg in convs:
                if msg.get("from") in ["human", "user"]:
                    user_msg = msg.get("value", "")
                    break
            if user_msg:
                eval_prompts.append({
                    "id": len(eval_prompts) + 1,
                    "prompt": user_msg
                })

    train_data = processed[:13500]
    val_data = processed[13500:15000]
    
    print(f"Writing {len(train_data)} train samples to {config.TRAIN_FILE}...")
    with open(config.TRAIN_FILE, "w", encoding="utf-8") as f:
        for item in train_data:
            f.write(json.dumps(item) + "\n")
            
    print(f"Writing {len(val_data)} validation samples to {config.VAL_FILE}...")
    with open(config.VAL_FILE, "w", encoding="utf-8") as f:
        for item in val_data:
            f.write(json.dumps(item) + "\n")
            
    print(f"Writing 100 holdout eval prompts to {config.EVAL_FILE}...")
    with open(config.EVAL_FILE, "w", encoding="utf-8") as f:
        json.dump(eval_prompts, f, indent=2)
        
    print("Dataset preparation complete!")

if __name__ == "__main__":
    main()