lm-evaluation-harness icon indicating copy to clipboard operation
lm-evaluation-harness copied to clipboard

chatglm2 acc=0 on lambada_openai dataset, is it correct?

Open changwangss opened this issue 2 years ago • 3 comments
trafficstars

I also validate chatglm and chatglm3, they are works, could you have fix the root cause? https://huggingface.co/THUDM/chatglm2-6b/discussions/97 python main.py --model hf-causal --model_args pretrained=THUDM/chatglm2-6b,trust_remote_code=True --tasks lambada_openai --limit 10 --batch_size 1 --no_cache

changwangss avatar Nov 02 '23 12:11 changwangss

When I run this I get AttributeError: property 'pad_token_id' of 'ChatGLMTokenizer' object has no setter. You're seeing it run though?

StellaAthena avatar Nov 02 '23 14:11 StellaAthena

as the code show, https://huggingface.co/THUDM/chatglm2-6b/blob/main/tokenization_chatglm.py#L91 yes, correct, the pad_token or pad_token_id is property and has no setter. I comment tokenizer.pad_token = tokenizer.eos_token, I can get the acc about chatglm&chatglm3, but acc=0 for chatglm2. Could you give me some help?

changwangss avatar Nov 09 '23 09:11 changwangss

ChatGLM-6B and ChatGLM3-6b both seem to work fine. I'm still unsure as to what's going wrong on ChatGLM2 and would very much like to figure out how to fix it.

haileyschoelkopf avatar Dec 22 '23 00:12 haileyschoelkopf