vllm-project/llama-68m

This is a copy of JackFram/llama-68m at commit 9de8453, made on 2026-10-30. vLLM CI uses this checkpoint and the original is on a personal account, so we keep a copy here to reduce supply-chain risk.

Model description

This is a LLaMA-like model with only 68M parameters trained on Wikipedia and part of the C4-en and C4-realnewslike datasets.

No evaluation has been conducted yet, so use it with care.

The model is mainly developed as a base Small Speculative Model in the SpecInfer paper.

Evaluations (contributed by Akshit, huge thanks!)

Category	Benchmark	Metric	Score / Value	Status
Linguistics & Grammar	BLiMP	Accuracy	70.57%	Success
Commonsense & Reasoning	PIQA	Normalized Accuracy	59.25%	Success
	BoolQ	Accuracy	57.71%	Success
	COPA	Accuracy	53.00%	Success
	WinoGrande	Accuracy	50.59%	Success
	HellaSwag	Normalized Accuracy	29.04%	Success
	RACE	Accuracy	25.36%	Success
	CommonsenseQA	Accuracy	19.82%	Success
Academic & Knowledge	SciQ	Normalized Accuracy	57.80%	Success
	ARC-Easy	Normalized Accuracy	35.98%	Success
	OpenBookQA	Normalized Accuracy	25.60%	Success
	MMLU	Accuracy	22.96%	Success
	ARC-Challenge	Normalized Accuracy	22.87%	Success
Language Modeling	TriviaQA	Accuracy	TriviaQA Standard	Success
	LAMBADA	Accuracy	13.24%	Success
	C4-Perplexity	Word Perplexity	205.79	Success
	WikiText-2	Word Perplexity	306.79	Success

Notes on Failed Tasks: The Arithmetic and SocialIQA benchmarks failed during execution due to runtime pipeline incompatibilities, yielding no score. Total evaluation runtime was 44.74 minutes.

Citation

To cite the model, please use

@misc{miao2023specinfer,
      title={SpecInfer: Accelerating Generative LLM Serving with Speculative Inference and Token Tree Verification}, 
      author={Xupeng Miao and Gabriele Oliaro and Zhihao Zhang and Xinhao Cheng and Zeyu Wang and Rae Ying Yee Wong and Zhuoming Chen and Daiyaan Arfeen and Reyna Abhyankar and Zhihao Jia},
      year={2023},
      eprint={2305.09781},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}