debug_dataset.py

#!/usr/bin/env python3
"""
Debug script to analyze the dataset structure.
"""

import json
from datasets import load_dataset
from collections import Counter

def analyze_sample(sample, idx):
    """Analyze a single sample structure."""
    conversations = sample.get('conversations', [])
    roles = [conv.get('from') or conv.get('role') for conv in conversations]

    has_function_call = 'function_call' in roles
    has_observation = 'observation' in roles

    return {
        'index': idx,
        'num_turns': len(conversations),
        'roles': roles,
        'has_function_call': has_function_call,
        'has_observation': has_observation,
        'has_tools': 'tools' in sample and sample['tools'] is not None
    }

def main():
    print("Loading dataset...")
    dataset = load_dataset("allenai/Dolci-Instruct-SFT-Tool-Use", split="train")

    print(f"Total samples: {len(dataset)}")

    # Analyze first few samples
    print("\n=== First 5 samples ===")
    for i in range(min(5, len(dataset))):
        info = analyze_sample(dataset[i], i)
        print(f"\nSample {i}:")
        print(f"  Roles: {info['roles']}")
        print(f"  Has function_call: {info['has_function_call']}")
        print(f"  Has observation: {info['has_observation']}")
        print(f"  Has tools: {info['has_tools']}")

    # Count role distribution
    print("\n=== Analyzing entire dataset ===")
    role_counter = Counter()
    tool_call_count = 0
    observation_count = 0
    both_count = 0

    for sample in dataset:
        conversations = sample.get('conversations', [])
        for conv in conversations:
            role = conv.get('from') or conv.get('role')
            role_counter[role] += 1

        roles = [conv.get('from') or conv.get('role') for conv in conversations]
        has_function = 'function_call' in roles
        has_observation = 'observation' in roles

        if has_function:
            tool_call_count += 1
        if has_observation:
            observation_count += 1
        if has_function and has_observation:
            both_count += 1

    print(f"\nRole distribution:")
    for role, count in role_counter.most_common():
        print(f"  {role}: {count}")

    print(f"\nTool calling statistics:")
    print(f"  Samples with function_call: {tool_call_count}")
    print(f"  Samples with observation: {observation_count}")
    print(f"  Samples with BOTH: {both_count}")

    # Check batch1 structure
    print("\n=== Checking batch1 structure ===")
    with open('data/dolci_10k_with_tool_call_batch1.json', 'r') as f:
        batch1 = json.load(f)

    print(f"Batch1 total samples: {len(batch1)}")
    if batch1:
        sample = batch1[0]
        print(f"Batch1 sample 0 keys: {sample.keys()}")
        print(f"Batch1 sample 0 conversations roles: {[c['from'] for c in sample['conversations']]}")
        print(f"First message value (first 100 chars): {sample['conversations'][0]['value'][:100]}")

if __name__ == "__main__":
    main()