MMLU-Pro category scores
#2
by
yaronr
- opened
Hi
We ran mmlu-pro on SaulLM-54B-Instruct, and the results below were quite surprising.
We expected SaulLM to ace the law category, but in fact, we saw only 30% correct answers.
Surprisingly, we saw the best mmlu-pro results for SaulLM-54-Instruct for psychology (62%) and biology (66%).
Appreciate your guidance.
We ran vllm latest (0.6.1.post2) with the following params:
--model=Equall/SaulLM-54B-Instruct
--tensor-parallel-size=2
--pipeline-parallel-size=4
--disable-log-requests
--trust-remote-code
--enable-chunked-prefill
--enable-prefix-caching
--enforce-eager
--max-num-batched-tokens=16384
--max-model-len=8096
--gpu-memory-utilization=0.95
Detailed results per category:
Thanks for the useful information, we have a legal LLM, we like to test, can you assist?