About model format conversion

[fzxxg]

I didn’t reproduce using your code, can I use your evaluation file? But I noticed that your model format needs to be converted. If I don’t convert it, it can still run successfully, but what impact will this have? The following is the code I executed for the evaluation file:
python evaluation.py
--model_name_or_path bert-base-uncased-sts
--pooler cls_before_pooler
--task_set sts/transfer
--mode test

[fzxxg]

I only saved the BERT parameters in the model.

[seanswyi]

I believe the output format is safetensors. There should be any actual impact.

[gaotianyu1350]

Hi @fzxxg , if you use our training code and don't convert it, the checkpoint won't be able to perform as well as the paper reported; however, if you are evaluating other models, it should be fine (as long as the inference for those models are strictly the same as how they are supposed to be used.

[fzxxg]

Thank you for your reply. I will close this issue.