Starting code to build parquets...

pdelfin/train/buildparquetdataset.py

@@ -0,0 +1,77 @@
+import argparse
+import logging
+from functools import partial
+import os
+import boto3
+from datasets import Dataset
+from botocore.exceptions import NoCredentialsError, PartialCredentialsError
+from pdelfin.train.dataloader import build_batch_query_response_vision_dataset
+
+
+def save_dataset_in_parquet(dataset: Dataset, output_dir: str, rows_per_file: int = 10000, s3_endpoint_url: str = None):
+    logger.info("Saving dataset in Parquet files")
+    
+    # Check if the output is an S3 path
+    is_s3 = output_dir.startswith("s3://")
+    if is_s3:
+        s3_client = boto3.client('s3', endpoint_url=s3_endpoint_url) if s3_endpoint_url else boto3.client('s3')
+    else:
+        os.makedirs(output_dir, exist_ok=True)
+    
+    total_rows = len(dataset)
+    for start_idx in range(0, total_rows, rows_per_file):
+        end_idx = min(start_idx + rows_per_file, total_rows)
+        file_name = f"dataset_{start_idx}_{end_idx}.parquet"
+        if is_s3:
+            # Saving to S3
+            bucket_name, key_prefix = parse_s3_path(output_dir)
+            output_path = f"{key_prefix}/{file_name}"
+            local_temp_file = f"/tmp/{file_name}"
+            logger.info(f"Saving rows {start_idx} to {end_idx} locally at {local_temp_file}")
+            dataset.select(range(start_idx, end_idx)).to_parquet(local_temp_file)
+            try:
+                logger.info(f"Uploading {local_temp_file} to s3://{bucket_name}/{output_path}")
+                s3_client.upload_file(local_temp_file, bucket_name, output_path)
+            except (NoCredentialsError, PartialCredentialsError) as e:
+                logger.error(f"Failed to upload to S3: {e}")
+                raise
+            finally:
+                os.remove(local_temp_file)
+        else:
+            # Saving locally
+            output_path = os.path.join(output_dir, file_name)
+            logger.info(f"Saving rows {start_idx} to {end_idx} in {output_path}")
+            dataset.select(range(start_idx, end_idx)).to_parquet(output_path)
+
+def parse_s3_path(s3_path: str):
+    """Parses an S3 path into bucket and key prefix."""
+    if not s3_path.startswith("s3://"):
+        raise ValueError("S3 path must start with 's3://'")
+    path = s3_path[5:]
+    bucket_name, _, key_prefix = path.partition('/')
+    return bucket_name, key_prefix
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Process and save dataset as Parquet files.")
+    parser.add_argument("--query_path", type=str, required=True, help="Path to the query dataset JSONL files.")
+    parser.add_argument("--response_path", type=str, required=True, help="Path to the response dataset JSONL files.")
+    parser.add_argument("--output_dir", type=str, required=True, help="Directory or S3 path to save the output Parquet files.")
+    parser.add_argument("--num_proc", type=int, default=32, help="Number of processes to use for data processing.")
+    parser.add_argument("--s3_endpoint_url", type=str, default=None, help="Custom S3 endpoint URL, e.g., for S3-compatible storage.")
+
+    args = parser.parse_args()
+
+    logging.basicConfig(level=logging.INFO)
+    logger = logging.getLogger(__name__)
+
+    # Build the dataset
+    final_dataset = build_batch_query_response_vision_dataset(
+        query_glob_path=args.query_path,
+        response_glob_path=args.response_path,
+        num_proc=args.num_proc
+    )
+
+    # Save the dataset as Parquet files
+    save_dataset_in_parquet(final_dataset, args.output_dir, s3_endpoint_url=args.s3_endpoint_url)
+
+    logger.info("Dataset processing and saving completed.")

pdelfin/train/core/config.py

@@ -75,8 +75,9 @@ class AwsConfig:
 @dataclass
 class SourceConfig:
     name: str = field(help="The name of the source")
-    query_glob_path: str = field(help="The s3 bucket pointing to the inputs sent to OpenAI to generate the silver data")
+    parquet_path: Optional[str] = field(help="The s3/glob path to a bunch of parquet files for a preprocessed dataset.", default=None)
-    response_glob_path: str = field(help="The s3 bucket pointing to the batch api response json's sent back from open ai")
+    query_glob_path: Optional[str] = field(help="The s3 bucket pointing to the inputs sent to OpenAI to generate the silver data", default=None)
+    response_glob_path: Optional[str] = field(help="The s3 bucket pointing to the batch api response json's sent back from open ai", default=None)
 
 
 @dataclass

pdelfin/train/dataloader.py

@@ -240,9 +240,6 @@ def build_batch_query_response_vision_dataset(query_glob_path: str, response_glo
 
     final_dataset = final_dataset.filter(pick_image_sizes, num_proc=num_proc)
 
-    # Limit the size of the input text not to explode the context size
-    final_dataset = final_dataset.filter(lambda x: len(x["raw_page_text"]) < 4000, num_proc=num_proc)
-
     return final_dataset