I am trying to use autotrain from Hugging Face to fine tune some models. Since I don’t have big computing resources, I tried to fine tune model EleutherAI/pythia-14m and this dataset. But I recieved this message:
Failed to merge adapter weights: Error(s) in loading state_dict for PeftModelForCausalLM:
size mismatch for base_model.model.gpt_neox.embed_in.weight: copying a param with shape torch.Size([50280, 128]) from checkpoint, the shape in current model is torch.Size([50277, 128]).
size mismatch for base_model.model.embed_out.weight: copying a param with shape torch.Size([50280, 128]) from checkpoint, the shape in current model is torch.Size([50277, 128]).
This error occured when I started this script, that is just terminal command, written in Jupiter.
!autotrain llm
--train
--model "EleutherAI/pythia-14m"
--project-name "my-llm"
--data-path data/
--text-column text
--batch-size "4"
--lr "2e-5"
--epochs "3"
--block-size "1024"
--warmup-ratio "0.03"
--lora-r "16"
--lora-alpha "32"
--lora-dropout "0.05"
--weight-decay "0."
--gradient-accumulation "4"
--logging-steps "10"
--use-peft
--merge-adapter
Also, the same problem showed, when I tried to autotrain in Hugging Face space.
I am not experienced in ml, so I can’t imagine, what can cause this problem.
New contributor