tokenizer.model_max_length=2048 in sample_finetune.py

#10
by anakin87 - opened

Hello and thx for the great work!

I noticed that in the fine-tuning example (sample_finetune.py)
you are setting tokenizer.model_max_length=2048 (https://huggingface.co/microsoft/Phi-3.5-mini-instruct/blob/64963004ad95869fa73a30279371c8778509ac84/sample_finetune.py#L140)

Based on my experience, this could be wrong.
The parameter max_seq_length=2048in SFTTrainer https://huggingface.co/microsoft/Phi-3.5-mini-instruct/blob/64963004ad95869fa73a30279371c8778509ac84/sample_finetune.py#L189
already takes care of truncating longer examples during fine-tuning.

Also, when setting tokenizer.model_max_length=2048, I experienced strange errors that prevented the fine-tuning process.

LMK what you think...

Sign up or log in to comment