Paper: https://arxiv.org/pdf/2310.06694.pdf
Code: https://github.com/princeton-nlp/LLM-Shearing
License: Must comply with license of Pythia since it's a model derived from Pythia.
Sheared-Pythia-160m is a model pruned and further pre-trained from EleutherAI/pythia-410m. We dynamically load data from different domains in the Pile dataset to prune and contune pre-train the model. We use 0.4B tokens for pruning and 50B tokens for continued pre-training the pruned model. This model can be loaded with HuggingFace via
model = GPTNeoXForCausalLM.from_pretrained("princeton-nlp/Sheared-Pythia-160m")
The model's overall performance is better than EleutherAI/pythia-160m.
Bibtex
@article{xia2023sheared,
title={Sheared llama: Accelerating language model pre-training via structured pruning},
author={Xia, Mengzhou and Gao, Tianyu and Zeng, Zhiyuan and Chen, Danqi},
journal={arXiv preprint arXiv:2310.06694},
year={2023}
}