attn-1b-50bn

Model Card

This model is pretrained as a reference baseline to the Based model provided here: https://huggingface.co/hazyresearch/based-1b-50b.

Both checkpoints are pretrained on 50Bn tokens of the Pile in the exact same data order using next token prediction.

Model Sources

The model is a standard Transformer model, using the Llama architecture (Rotary encodings, SwiGLU, RMS Norm, etc.)

The training code is provided here and can be used to reproduce training: https://github.com/HazyResearch/based

The paper for the work is here, and the appendix includes additional experimental details/hyperparameters: https://arxiv.org/abs/2402.18668

Uses

The purpose of this work is to evaluate the language modeling quality of a new efficient architecture, Based.

We include a series of benchmarks that you can use to evaluate quality:

• FDA: https://huggingface.co/datasets/hazyresearch/based-fda
• SWDE: https://huggingface.co/datasets/hazyresearch/based-swde
• SQUAD: https://huggingface.co/datasets/hazyresearch/based-squad

Citation

Please consider citing this paper if you use our work:

@article{arora2024simple,
  title={Simple linear attention language models balance the recall-throughput tradeoff},
  author={Arora, Simran and Eyuboglu, Sabri and Zhang, Michael and Timalsina, Aman and Alberti, Silas and Zinsley, Dylan and Zou, James and Rudra, Atri and Ré, Christopher},
  journal={arXiv:2402.18668},
  year={2024}
}

Please reach out to [email protected], [email protected], and [email protected] with questions.