metadata

license: apache-2.0

Paper: https://arxiv.org/pdf/2310.06694.pdf
Code: https://github.com/princeton-nlp/LLM-Shearing
Models: Sheared-LLaMA-1.3B, Sheared-LLaMA-2.7B
Instruction-tuned Models: Sheared-LLaMA-1.3B-ShareGPT, Sheared-LLaMA-2.7B-ShareGPT

License: Must comply with license of Llama2 since it's a model derived from Llama2.

Sheared-LLaMA-1.3B is a model pruned and further pre-trained from meta-llama/Llama-2-7b-hf. We dynamically load data from different domains in the RedPajama 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 = AutoModelForCausalLM.from_pretrained("princeton-nlp/Sheared-LLaMA-1.3B")

Smaller-scale
Same vocabulary as LLaMA1 and LLaMA2
Derived with a budget of 50B tokens by utilizing existing strong LLMs

Downstream Tasks

We evaluate on an extensive set of downstream tasks including reasoning, reading comprehension, language modeling and knowledge intensive tasks. Our Sheared-LLaMA models outperform existing large language models.

Model	# Pre-training Tokens	Average Performance
LLaMA2-7B	2T	64.6

1.3B

Model	# Pre-training Tokens	Average Performance
OPT-1.3B	300B	48.2
Pythia-1.4B	300B	48.9
Sheared-LLaMA-1.3B	50B	51.0

Model	# Pre-training Tokens	Average Performance
OPT-2.7B	300B	51.4
Pythia-2.8B	300B	52.5
INCITE-Base-3B	800B	54.7
Open-LLaMA-3B-v1	1T	55.1
Open-LLaMA-3B-v2	1T	55.7
Sheared-LLaMA-2.7B	50B	56.7

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}
}

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	31.47
ARC (25-shot)	32.85
HellaSwag (10-shot)	60.91
MMLU (5-shot)	25.71
TruthfulQA (0-shot)	37.14
Winogrande (5-shot)	58.64
GSM8K (5-shot)	0.45
DROP (3-shot)	4.56