preprocess.py

import json
from data_loader import get_combined_verilog_data

def format_for_prolog_agent(example):
    """
    Transforms a raw Verilog dataset entry into a 
    Training Prompt for Agent 1.
    """
    instruction = example['instruction']
    
    # SYSTEM PROMPT: Tells the LLM how to act
    system_message = (
        "You are a Hardware Verification Agent. "
        "Translate the following hardware description into formal Prolog constraints."
    )
    
    # FORMAT: This is the structure the LLM will learn
    # User: English Text -> Assistant: Prolog Code
    formatted_text = f"### System: {system_message}\n"
    formatted_text += f"### Instruction: {instruction}\n"
    formatted_text += f"### Response: [PROLOG_START]\n" 
    # Note: In a real training set, you'd put your manual Prolog here as the 'target'
    
    return {"text": formatted_text}

def prepare_training_file(output_file="training_data.jsonl"):
    print("Loading datasets for preprocessing...")
    dataset = get_combined_verilog_data()
    
    print(f"Processing {len(dataset)} entries...")
    processed_data = dataset.map(format_for_prolog_agent)
    
    # Save to JSONL (standard format for Hugging Face fine-tuning)
    with open(output_file, "w") as f:
        for entry in processed_data:
            f.write(json.dumps(entry) + "\n")
            
    print(f"✅ Preprocessing complete. Data saved to {output_file}")

if __name__ == "__main__":
    prepare_training_file()