Llemma: an open language model for mathematics

ArXiv | Models | Data | Code | Blog | Sample Explorer

Repository for Llemma: an open language model for mathematics [Azerbayev et al 2023].

This repository hosts data and training code related to the following artifacts:

Name	HF Hub Link
Llemma 7b	EleutherAI/llemma_7b
Llemma 34b	EleutherAI/llemma_34b
Proof-Pile-2	EleutherAI/ProofPile2
AlgebraicStack	EleutherAI/AlgebraicStack

This repository also contains submodules related to the overlap, fine-tuning, and theorem proving experiments described in the paper. Additional evaluation code is in a fork of the Eleuther LM Evaluation Harness.

Directories

This repository contains the following directories

• proof_pile_2: scripts for downloading and preprocessing data.
• gpt-neox: git submodule containing a modified branch of EleutherAI/gpt-neox
• lm-evaluation-harness: code for all evaluations, except formal2formal theorem proving.
• llemma_formal2formal: git submodule containing scripts for the formal2formal experiments
• overlap: git submodule containing the overlap and memorization analysis
• finetunes: git submodule containing scripts for the fine-tuning experiments

Because this project contains submodules, you should clone this project with the --recurse-submodules flag or, alternatively, run git submodule update --init --recursive from within the project directory after cloning the project. After running git pull, you should also run git submodule update.

Citation

Please cite the following:

@article{azerbayev2023llemma,
  title={Llemma: An Open Language Model For Mathematics}, 
  author={Azerbayev, Zhangir and Schoelkopf, Hailey and Paster, Keiran and Dos Santos, Marco and McAleer, Stephen and Jiang, Albert Q. and Deng, Jia and Biderman, Stella and Welleck, Sean},
  journal={arXiv preprint arXiv:2310.06786},
  year={2023}
}

You may also be interested in citing our training data, which is a mix of novel data and data from the following sources:

@article{paster2023openwebmath,
  title={OpenWebMath: An Open Dataset of High-Quality Mathematical Web Text},
  author={Paster, Keiran and Santos, Marco Dos and Azerbayev, Zhangir and Ba, Jimmy},
  journal={arXiv preprint arXiv:2310.06786},
  year={2023}
}

@software{together2023redpajama,
  author = {Together Computer},
  title = {RedPajama: An Open Source Recipe to Reproduce LLaMA training dataset},
  month = April,
  year = 2023,
  url = {https://github.com/togethercomputer/RedPajama-Data}
}

@article{kocetkov2022stack,
  title={The stack: 3 tb of permissively licensed source code},
  author={Kocetkov, Denis and Li, Raymond and Allal, Loubna Ben and Li, Jia and Mou, Chenghao and Ferrandis, Carlos Mu{\~n}oz and Jernite, Yacine and Mitchell, Margaret and Hughes, Sean and Wolf, Thomas and Bahdanau, Dzmitry and von Werra, Leandro and de Vries, Harm},
  journal={arXiv preprint arXiv:2211.15533},
  year={2022}
}