Closed starjob42 closed 1 month ago
Thanks for the feedback. I think that's because this model is the 4k model. There is a 128k context length Phi 3.5 model, but that's not supported yet since it would require LongRoPE, which is not implemented, yet.
Hi,
Thank you for awesome library!
I am using
litgpt version 0.4.11
Currently I am usingPhi-3.5-mini-instruct
to finetune using lora. Even though I set--train.max_seq_length 10000
I still getHere is the command that I am using to run:
litgpt finetune_lora microsoft/Phi-3.5-mini-instruct --data JSON --data.json_path my_data.json --data.val_split_fraction 0.1 --out_dir out/custom-model --train.max_seq_length 10000 --precision bf16-true --quantize bnb.nf4-dq # execute the litgpt command