Baichuan-13B icon indicating copy to clipboard operation
Baichuan-13B copied to clipboard

baichuan2 mmlu结果复现的问题

Open zhanghan1992 opened this issue 1 year ago • 0 comments

评估使用的代码:https://github.com/baichuan-inc/Baichuan-7B/blob/main/evaluation/evaluate_mmlu.py

用bf16精度测试 llama2-13-hf 和 baichuan2-13b-base llama2-13-hf: 0.550 baichuan2-13b-base: 0.564

改了一行代码,用fp32测试: #model = AutoModelForCausalLM.from_pretrained(args.model, torch_dtype=torch.bfloat16, device_map="auto",trust_remote_code=True) model = AutoModelForCausalLM.from_pretrained(args.model, device_map="auto",trust_remote_code=True) llama2-13-hf: 0.554 baichuan2-13b-base: 0.590

请教下,为啥baichuan2在bf16和fp32精度下结果差这么多?

zhanghan1992 avatar Nov 02 '23 09:11 zhanghan1992