UltraData-Stack-v3-L2

📦 UltraData Collection | 🌐 UltraData | 📖 Tech Report | 🤗 UltraData-Code | 🤗 UltraData-Code-L2 Classifier

English | 中文

📚 Introduction

UltraData-Stack-v3-L2 is a selected code corpus of approximately 240B tokens for the 11 programming languages in The Stack v3, obtained through quality-control cleaning and fine-grained selection with UltraData-Code-L2-Classifier. This dataset transfers the language-adaptive L2 selection framework proposed by UltraData-Code to another source corpus, in order to examine whether file-level algorithm-related relevance selection can generalize to data sources beyond UltraData-Code-L1.

Construction has two stages. First, source records are cleaned by removing URL-only entries, overly short records, and records containing garbled or malformed text. Second, a language-specific selector combines file-role prediction, algorithmic-relevance scoring, an EXCLUDE veto, and a shared code-quality constraint. The relevance threshold is calibrated separately for each language. The resulting dataset retains 20.10% of the cleaned 11-language candidate pool, and is intended to concentrate interpretable computation and problem-solving logic while preserving broad language coverage.

📢 What's New

• [2026.10.16] The UltraData-Code technical report is released! 🚀🚀🚀
• [2026.10.12] The UltraData-Stack-v3-L2 dataset is released! 🚀🚀🚀
• [2026.09.07] The UltraData-Code dataset is released! It is a complete implementation of the UltraData L0-L4 tiered data management framework. It covers four code data states from L0 through L3, with each level corresponding to a distinct construction stage. This release currently open-sources L2 (~400B tokens) and L3 (~150B tokens) across 11 programming languages. 🚀🚀🚀
• [2026.09.07] MiniCPM5-2B is released!, the second model in the MiniCPM5 series after MiniCPM5-1B. It is a dense 2B Transformer that scales up the same training recipe, built for on-device, local deployment, and resource-constrained scenarios. It reaches 2B-class open-source SOTA, remains competitive with 4B-class models, and shows particular advantages in coding, mathematics, long-context understanding, tool use, and agentic tasks. UltraData-RL-2609 serves as the core RL dataset for MiniCPM5-2B. 🚀🚀🚀
• [2026.02.08] The UltraData platform is now live, introducing the L0-L4 tiered data management framework. 🔍🔍🔍

💡 Highlights

• Transferable algorithmic-relevance selection: The same selection objective and model interfaces used for UltraData-Code-L2 are adapted to Stack-v3-train, with language-specific evidence and thresholds.
• Strong performance gains: Under the same 1B base model and 10B-token continual-pretraining setup, UltraData-Stack-v3-L2 reaches 30.00 average pass@1 on EvalPlus and 22.75 on MultiPL-E, compared with 25.37 and 17.30 for Stack-v3-train.

📊 Dataset Statistics

Token counts are estimated with the tokenizer used in the technical report.

Language	After cleaning	After selection	Retention	Token share	Tokens
Python	96,961,008	24,671,724	25.44%	15.86%	38.94B
Java	165,346,828	41,963,467	25.38%	28.76%	70.60B
JavaScript	110,967,790	14,521,755	13.09%	12.92%	31.73B
C++	67,981,897	17,990,102	26.46%	12.78%	31.38B
PHP	65,565,149	5,500,141	8.39%	10.26%	25.19B
Ruby	32,667,620	6,389,441	19.56%	2.39%	5.86B
Shell	26,704,436	5,045,813	18.90%	8.83%	21.67B
C#	17,309,141	3,548,274	20.50%	3.64%	8.93B
Go	29,683,899	3,138,105	10.57%	1.65%	4.04B
Rust	11,674,463	2,907,231	24.90%	2.13%	5.24B
R	7,640,182	1,456,974	19.07%	0.78%	1.92B
Total	632,502,413	127,133,027	20.10%	100%	245.50B

📈 Evaluation Results

The 10B-token comparisons use the same 1B foundation model, training settings, decontamination, and evaluation protocol.

Python results: UltraData-Stack-v3-L2-py ranks second among the evaluated L2 datasets by average EvalPlus pass@1 and obtains the highest MBPP and MBPP+ scores at L2.

Multilingual results: UltraData-Stack-v3-L2 exceeds Stack-v3-train by 4.63 points on EvalPlus and 5.45 points on MultiPL-E, respectively, demonstrating effective L2 language-adaptive selection.

❤️ Acknowledgements

UltraData-Stack-v3-L2 is built from The Stack v3. The selection framework UltraData-Code-L2-Classifier is developed as part of UltraData-Code, and its file-level representations are computed with Qwen3-Embedding-0.6B.

Thanks to the authors and contributors of these open-source projects.

💳 License

The dataset is released under the Open Data Commons Attribution License (ODC-By) v1.0, consistent with the source dataset The Stack v3. Any use of all or part of the code gathered in UltraData-Stack-v3-L2 must abide by the terms of the original licenses, including attribution clauses where relevant.

📖 Citation

If you find this dataset useful in your research, please consider citing:

@article{tu2026ultradata,
  title={{UltraData-Code}: From Raw Repositories to Algorithmically Dense and Task-Oriented Code Data},
  author={Tu, Chengying and Zhao, Hengyu and Xue, Shuaikang and Qu, Zhongming and Zhou, Jihao and Lin, Xinle and Guo, Junshao and Fu, Zixuan and Ma, Qiang and Zhou, Jie and others},
  journal={arXiv preprint arXiv:2610.23960},
  year={2026}
}