projects/elizabeth/training/create_clean_dataset.py

#!/usr/bin/env python3
"""
Create clean training dataset with only valid examples
"""

import json

def create_clean_dataset(input_path, output_path):
    """Create a clean dataset with only valid text examples"""
    clean_data = []
    
    with open(input_path, 'r') as f:
        for line in f:
            try:
                data = json.loads(line)
                
                # Only keep examples with simple text (no tool calls)
                messages = data.get('messages', [])
                
                # Skip if any message contains tool_call (complex structure)
                has_tool_call = any('tool_call' in str(msg) for msg in messages)
                if has_tool_call:
                    continue
                    
                # Only keep if we have at least 2 messages
                if len(messages) >= 2:
                    # Ensure each message has required fields
                    valid_messages = []
                    for msg in messages:
                        if isinstance(msg, dict) and 'role' in msg and 'content' in msg:
                            valid_messages.append({
                                'role': msg['role'],
                                'content': str(msg['content'])
                            })
                    
                    if len(valid_messages) >= 2:
                        clean_data.append({'messages': valid_messages})
                        
            except json.JSONDecodeError:
                continue
    
    # Write clean data
    with open(output_path, 'w') as f:
        for entry in clean_data:
            f.write(json.dumps(entry, ensure_ascii=False) + '\n')
    
    print(f"✅ Created clean dataset with {len(clean_data)} examples")
    print(f"💾 Saved to: {output_path}")
    return clean_data

# Create clean datasets
clean_train = create_clean_dataset(
    "/home/x/adaptai/aiml/e-train-1/elizabeth_tooluse_minipack_v1.jsonl",
    "/home/x/adaptai/aiml/e-train-1/clean_training_data.jsonl"
)

print(f"Total clean examples: {len(clean_train)}")