AI Potluck
Contribute
Back to Gap Map Model components / Training & synthetic datasets

Stack-Edu

Hugging Face
open / Overall score: 3.5

Hugging Face's education-filtered slice of StarCoder2Data, about 125 billion tokens of code scored by an educational-value classifier. It is a documented component of SmolLM3 and of the Dolma 3 pool behind OLMo 3.

Openness

4 medium confidence
4.0
access
ungated
license
inherits-the-stack-v2
dataset_card
present

Ungated download with a card; the card defers the data license to The Stack v2's terms.

  • https://huggingface.co/datasets/HuggingFaceTB/stack-edu recorded 2026-09-29

    The dataset page renders the card, which says "Stack-Edu is a 125B token dataset of educational code filtered from" The Stack v2 and "It is intended for Language Models training.", and its embedded repo data reports gated false and private false.

  • https://huggingface.co/datasets/HuggingFaceTB/stack-edu/raw/main/README.md recorded 2026-09-29

    The raw card says "Please refer to [the-stack-v2](https://huggingface.co/datasets/bigcode/the-stack-v2-train-full-ids) for the data license" and describes "Stack-Edu is a 125B token dataset of educational code filtered from" The Stack v2.

Adoption

2 high confidence
2.0

Adoption is measured by Hugging Face downloads of the dataset over the trailing month, the only usage channel on record for Stack-Edu.

Capability

5 high confidence
5.0

Stack-Edu beat StarCoder2Data on MultiPL-E in the SmolLM2 paper's annealing ablations, and the dataset card reports the gain across all its programming languages. The SmolLM3 pretraining mixture adds it from its second stage, and it is one of the code sources in the Dolma 3 pool behind Olmo 3.

Verified 2026-09-27