The output returning input.

[TemaLykov]

I finetuned the decapoda and yahma versions of thr alpaca on my custom dataset, but I always got the same result - the generated output looks like: ###Instruction... and so on, repeating my input. I saw issues with the same problem and i see that problem can be with the eos and pad tokens after changing tokenization format. But if i got it right, using last version of tloen/alpaca-lora with: model.config.pad_token_id = tokenizer.pad_token_id = 0 # unk model.config.bos_token_id = 1 model.config.eos_token_id = 2 should have solve the problem. I also used yahma/llama-7b-hf. I saw in one issue that can resolve it. But all my attempts finishes the same way - the repeating of my input. what I misunderstand of doing wrong?

[rogerbarretocode]

i am facing similar issue while finetuning with decapoda although it gives me proper output on some inputs

[DanqingZ]

Hi, you can try my script to test if your model is performing ok: https://github.com/tloen/alpaca-lora/issues/454

I also have inputs in my response when I use the generate.py file, but I did some debugging and it looks like the trained model is performing ok.