metadata
license: mit
Please refer to the SepLLM paper - ICML 2025 and our GitHub repository for using this model.
To use the checkpoint of this model, you must install the transformers-4.38.0.post1+sepllm-py3-none-any.whl released from our GitHub repository. Below are the reference script for testing and a sample of test results. We conducted testing using lm_eval==0.4.0.
CUDA_LAUNCH_BLOCKING=1
lm_eval --model hf \
--model_args pretrained=Gausson/pythia-160m-deduped-n128-SepLLM \
--tasks arc_challenge,arc_easy,lambada_openai,logiqa,piqa,sciq,winogrande,wsc,wikitext \
--num_fewshot 5 \
--device cuda:0\
--batch_size 32
hf (pretrained=Gausson/pythia-160m-deduped-n128-SepLLM), gen_kwargs: (), limit: None, num_fewshot: 5, batch_size: 32
| Tasks |Version|Filter|n-shot| Metric | Value | |Stderr|
|--------------|-------|------|-----:|---------------|------:|---|-----:|
|arc_challenge |Yaml |none | 5|acc | 0.2014|± |0.0117|
| | |none | 5|acc_norm | 0.2346|± |0.0124|
|arc_easy |Yaml |none | 5|acc | 0.4731|± |0.0102|
| | |none | 5|acc_norm | 0.4520|± |0.0102|
|lambada_openai|Yaml |none | 5|perplexity |30.1605|± |1.0128|
| | |none | 5|acc | 0.3315|± |0.0066|
|logiqa |Yaml |none | 5|acc | 0.2273|± |0.0164|
| | |none | 5|acc_norm | 0.2857|± |0.0177|
|piqa |Yaml |none | 5|acc | 0.6464|± |0.0112|
| | |none | 5|acc_norm | 0.6447|± |0.0112|
|sciq |Yaml |none | 5|acc | 0.8260|± |0.0120|
| | |none | 5|acc_norm | 0.8150|± |0.0123|
|wikitext |Yaml |none | 5|word_perplexity|30.3488| | |
| | |none | 5|byte_perplexity| 1.8931| | |
| | |none | 5|bits_per_byte | 0.9207| | |
|winogrande |Yaml |none | 5|acc | 0.5178|± |0.0140|
|wsc |Yaml |none | 5|acc | 0.3750|± |0.0477|
If you find our work helpful, please consider giving us a star ⭐ @ our GitHub repository and citing our paper. We greatly appreciate your support 😄
@inproceedings{chen2025sepllm,
title={{SepLLM: Accelerate Large Language Models by Compressing One Segment into One Separator}},
author={Chen, Guoxuan and Shi, Han and Li, Jiawei and Gao, Yihang and Ren, Xiaozhe and Chen, Yimeng and Jiang, Xin and Li, Zhenguo and Liu, Weiyang and Huang, Chao},
booktitle={Proceedings of the Forty-Second International Conference on Machine Learning (ICML)},
year={2025},
note={Also available at arXiv:2412.12094}
}