💻 Stack-Edu

Stack-Edu is a 125B token dataset of educational code filtered from The Stack v2, precisely the curated training corpus of StarCoder2 models denoted StarCoder2Data. It is intended for Language Models training.

This dataset was curated using a classifier-based filtering strategy, inspired by 📚 FineWeb-Edu, to retain only the highest-quality educational programming content.

Stack-Edu shows consistent improvement over StarCoder2data on all the programming languages on MultiPL-E benchmark.

Downloading the data

This dataset only contains the SWHIDs to download the code files and not the content of the files itself. The contents can be downloaded from Software Heritage's S3 bucket to ensure data compliance. Please refer to the-stack-v2 for the data license.

When running on a 16-core AWS us-east-1 instance, this script takes ~6 hours to download the files:

import boto3
import gzip
from datasets import load_dataset
from botocore.exceptions import ClientError

num_proc = 16
s3 = boto3.client('s3')
bucket_name = "softwareheritage"

def download_contents(blob_id):
    key = f"content/{blob_id}"
    try:
        obj = s3.get_object(Bucket=bucket_name, Key=key)
        with gzip.GzipFile(fileobj=obj['Body']) as fin:
            content = fin.read().decode("utf-8", errors="ignore")
        return {"text": content, "download_success": True}
    except ClientError as e:
        if e.response['Error']['Code'] == 'NoSuchKey':
            print(f"File not found: {key}")
            return {"text": "", "download_success": False}
        else:
            raise

# For Python
ds = load_dataset("HuggingFaceTB/stack-edu", "Python", split="train", num_proc=num_proc)
ds = ds.map(download_contents, input_columns="blob_id", num_proc=num_proc)

# Filter out failed downloads
ds = ds.filter(lambda x: x['download_success'])

# Optionally, print the first example to verify the data
print(ds[0])

Details

The table below shows the number of tokens in each programming language using SmolLM2 tokenizer.

Language	Stack-Edu (B tokens)
Python	21.8
Cpp	16.0
Markdown	14.0
C	11.1
JavaScript	11.1
Java	42.1
SQL	9.62
PHP	9.07
C-Sharp	8.87
TypeScript	3.03
Shell	3.13
Swift	1.83
Go	1.80
Rust	1.75
Ruby	1.61

Dataset curation

To build Stack-Edu, we:

• Selected 15 largest programming languages from StarCoder2Data
• Trained 15 language-specific classifiers, using StarEncoder model on synthetic annotations generated by Llama3-70B-Instruct. The classifiers for each language are available in this collection.
• Applied a filtering threshold of 3 (out of 5) to retain highly educational content, except for Java, which performed best at threshold 2.

Citation Information

@misc{allal2025smollm2smolgoesbig,
      title={SmolLM2: When Smol Goes Big -- Data-Centric Training of a Small Language Model}, 
      author={Loubna Ben Allal and Anton Lozhkov and Elie Bakouch and Gabriel Martín Blázquez and Guilherme Penedo and Lewis Tunstall and Andrés Marafioti and Hynek Kydlíček and Agustín Piqueres Lajarín and Vaibhav Srivastav and Joshua Lochner and Caleb Fahlgren and Xuan-Son Nguyen and Clémentine Fourrier and Ben Burtenshaw and Hugo Larcher and Haojun Zhao and Cyril Zakka and Mathieu Morlon and Colin Raffel and Leandro von Werra and Thomas Wolf},
      year={2025},
      eprint={2502.02737},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2502.02737}, 
}