---
license: mit
---

```

import numpy as np
import json
import os
from huggingface_hub import snapshot_download
from tqdm import tqdm

# ═══════════════════════════════════════════════════════════
#  CONFIG
# ═══════════════════════════════════════════════════════════

HF_REPO = "nnsohamnn/runner-game-dataset"  # ← Change this!
DOWNLOAD_DIR = "runner_dataset_merged"
OUTPUT_DIR = "runner_dataset"

# ═══════════════════════════════════════════════════════════
#  DOWNLOAD
# ═══════════════════════════════════════════════════════════

print("📥 Downloading from Hugging Face...")
snapshot_download(
    repo_id=HF_REPO,
    repo_type="dataset",
    local_dir=DOWNLOAD_DIR
)
print("✅ Download complete!")

# ═══════════════════════════════════════════════════════════
#  OPTION A: USE MERGED FILES DIRECTLY (RECOMMENDED)
# ═══════════════════════════════════════════════════════════

# You can use the merged files directly in training!
# This is actually MORE efficient than individual files.

# Example loading:
print("\n📊 Dataset info:")
with open(os.path.join(DOWNLOAD_DIR, "metadata.json"), 'r') as f:
    metadata = json.load(f)
    print(f"   Total frames: {metadata['total_frames']:,}")
    print(f"   Chunks: {metadata['num_chunks']}")
    print(f"   Actions: {metadata['actions']}")

# ═══════════════════════════════════════════════════════════
#  OPTION B: UNPACK TO INDIVIDUAL FILES (if needed)
# ═══════════════════════════════════════════════════════════

def unpack_dataset():
    """Unpack merged files back to individual files (optional)"""
    
    print("\n📦 Unpacking to individual files...")
    
    os.makedirs(os.path.join(OUTPUT_DIR, "frames"), exist_ok=True)
    os.makedirs(os.path.join(OUTPUT_DIR, "actions"), exist_ok=True)
    
    # Unpack frames
    chunk_files = sorted([f for f in os.listdir(DOWNLOAD_DIR) if f.startswith("frames_chunk")])
    
    frame_idx = 0
    for chunk_file in chunk_files:
        print(f"   Unpacking {chunk_file}...")
        data = np.load(os.path.join(DOWNLOAD_DIR, chunk_file))
        frames = data['frames']
        
        for i in tqdm(range(len(frames)), desc=f"   {chunk_file}"):
            np.save(
                os.path.join(OUTPUT_DIR, "frames", f"{frame_idx:06d}.npy"),
                frames[i]
            )
            frame_idx += 1
        
        data.close()
    
    # Unpack actions
    print("   Unpacking actions.jsonl...")
    with open(os.path.join(DOWNLOAD_DIR, "actions.jsonl"), 'r') as f:
        for idx, line in enumerate(tqdm(f, desc="   actions")):
            action_data = json.loads(line)
            with open(os.path.join(OUTPUT_DIR, "actions", f"{idx:06d}.json"), 'w') as out_f:
                json.dump(action_data, out_f)
    
    print(f"\n✅ Unpacked to {OUTPUT_DIR}/")

# Uncomment to unpack:
# unpack_dataset()

```