Datasets:

togethercomputer

/

RedPajama-Data-V2

like 308

Thanks to Together team for preparing RPV2 as well as the sample 1B dataset!!

I have a need to download the sample data locally to debug and test processing scripts in Apache Beam. load_dataset("togethercomputer/RedPajama-Data-V2", name="sample") download files as arrow instead of the original formats, which is not good for my case.

So I made a script to download all sample files locally via hf_hub_download. Thought it might be helpful for others who have similar needs.

import os
from huggingface_hub import hf_hub_download


repo_id = "togethercomputer/RedPajama-Data-V2"
sample_list_name = "sample/sample_listings.txt"

direcotories_postfix_map = {
    "documents": ".json.gz",
    "quality_signals": ".signals.json.gz",
    "minhash": ".minhash.parquet",
    "duplicates": ".duplicates.parquet",
}

# download and read sample_listings.txt
filepath = hf_hub_download(repo_id=repo_id, filename=sample_list_name, repo_type="dataset")
with open(filepath, "r") as f:
    sample_listings = f.read().splitlines()

# download all files in sample_listings.txt
for sample_listing in sample_listings:
    for directory, postfix in direcotories_postfix_map.items():
        filename = f"sample/{directory}/{sample_listing}{postfix}"
        cache_file_path = hf_hub_download(
            repo_id=repo_id,
            filename=filename,
            repo_type="dataset",
        )
        print(f"Downloaded {filename} to {cache_file_path}")