dataset_info:
features:
- name: blob_id
dtype: string
- name: repo_name
dtype: string
- name: path
dtype: string
- name: length_bytes
dtype: int64
- name: score
dtype: float64
- name: int_score
dtype: int64
- name: text
dtype: string
- name: download_success
dtype: bool
splits:
- name: train
num_bytes: 13499266964
num_examples: 7678448
download_size: 6086016638
dataset_size: 13499266964
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
About This Dataset
The HuggingFaceTB team has released an impressive series of models called smollm (V1/V2) (paper： https://arxiv.org/abs/2502.02737).
According to their documentation, they used Stack-Edu as the code field corpus for pretraining and published https://huggingface.co/datasets/HuggingFaceTB/smollm-corpus.
However for some reason, only a Python-Edu subset is accessible and there's no content/text field in it.
The full dataset is stored on AWS S3; downloading it requires an AWS EC2 instance, or you will be blocked by AWS's rate limits under which you will never download it.
Fortunately, the py-edu subset is relatively small (~7 million files) for me to afford; downloading the entire set takes approximately one hour.
I am publishing the complete py-edu dataset here for anyone who needs it.
If this release inadvertently causes any issues for the HuggingFaceTB team, please reach out to me and I will remove it immediately.