Add automatic dataset sampling to reduce disk usage

When --sample is specified, pre-sample dataset using streaming to avoid
storing full dataset on disk. Reduces space usage from GBs to MBs.

🤖 Generated with [Claude Code](https://claude.ai/code)

Co-Authored-By: Claude <[email protected]>

atlas-export.py

@@ -3,6 +3,7 @@
 # dependencies = [
 #     "huggingface-hub[hf_transfer]",
 #     "torch",  # For GPU detection
+#     "datasets",  # For dataset sampling
 # ]
 # ///
 
@@ -71,8 +72,58 @@ def check_gpu_available() -> bool:
         return False
 
 
-def build_atlas_command(args) -> list:
+def sample_dataset_to_parquet(
+    dataset_id: str, 
+    sample_size: int, 
+    split: str = "train",
+    trust_remote_code: bool = False
+) -> tuple[Path, Path]:
+    """Sample dataset and save to local parquet file."""
+    from datasets import load_dataset
+    
+    logger.info(f"Pre-sampling {sample_size} examples from {dataset_id}...")
+    
+    # Load with streaming to avoid loading entire dataset
+    ds = load_dataset(
+        dataset_id, 
+        streaming=True, 
+        split=split,
+        trust_remote_code=trust_remote_code
+    )
+    
+    # Shuffle and sample
+    ds = ds.shuffle(seed=42)
+    sampled_ds = ds.take(sample_size)
+    
+    # Create temporary directory and parquet file
+    temp_dir = Path(tempfile.mkdtemp(prefix="atlas_data_"))
+    parquet_path = temp_dir / "data.parquet"
+    
+    logger.info(f"Saving sampled data to temporary file...")
+    sampled_ds.to_parquet(str(parquet_path))
+    
+    file_size = parquet_path.stat().st_size / (1024 * 1024)  # MB
+    logger.info(f"Created {file_size:.1f}MB parquet file with {sample_size} samples")
+    
+    return parquet_path, temp_dir
+
+
+def build_atlas_command(args) -> tuple[list, str, Optional[Path]]:
     """Build the embedding-atlas command with all parameters."""
+    temp_data_dir = None
+    
+    # If sampling is requested, pre-sample the dataset
+    if args.sample:
+        parquet_path, temp_data_dir = sample_dataset_to_parquet(
+            args.dataset_id,
+            args.sample,
+            args.split,
+            args.trust_remote_code
+        )
+        dataset_input = str(parquet_path)
+    else:
+        dataset_input = args.dataset_id
+    
     # Use uvx to run embedding-atlas with required dependencies
     # Include hf-transfer for faster downloads when HF_HUB_ENABLE_HF_TRANSFER is set
     cmd = [
@@ -82,7 +133,7 @@ def build_atlas_command(args) -> list:
         "--with",
         "hf-transfer",
         "embedding-atlas",
-        args.dataset_id,
+        dataset_input,
     ]
     # Add all optional parameters
     if args.model:
@@ -95,11 +146,12 @@ def build_atlas_command(args) -> list:
     if args.image_column:
         cmd.extend(["--image", args.image_column])
 
-    if args.split:
+    if args.split and not args.sample:  # Don't add split if we pre-sampled
         cmd.extend(["--split", args.split])
 
-    if args.sample:
-        cmd.extend(["--sample", str(args.sample)])
+    # Remove the --sample parameter since we've already sampled
+    # if args.sample:
+    #     cmd.extend(["--sample", str(args.sample)])
 
     if args.trust_remote_code:
         cmd.append("--trust-remote-code")
@@ -120,7 +172,7 @@ def build_atlas_command(args) -> list:
     export_path = "atlas_export.zip"
     cmd.extend(["--export-application", export_path])
 
-    return cmd, export_path
+    return cmd, export_path, temp_data_dir
 
 
 def create_space_readme(args) -> str:
@@ -399,9 +451,11 @@ def main():
         output_dir = Path(temp_dir)
         logger.info(f"Using temporary directory: {output_dir}")
 
+    temp_data_dir = None  # Initialize to avoid UnboundLocalError
+    
     try:
         # Build and run embedding-atlas command
-        cmd, export_path = build_atlas_command(args)
+        cmd, export_path, temp_data_dir = build_atlas_command(args)
         logger.info(f"Running command: {' '.join(cmd)}")
 
         # Run the command
@@ -445,6 +499,11 @@ def main():
         if temp_dir and not args.local_only:
             shutil.rmtree(temp_dir)
             logger.info("Cleaned up temporary files")
+        
+        # Also clean up the data sampling directory
+        if temp_data_dir and temp_data_dir.exists():
+            shutil.rmtree(temp_data_dir)
+            logger.info("Cleaned up sampled data")
 
 
 if __name__ == "__main__":