About the max length of the "PMC_LLaMA_13B"

[dyxohjl666]

Hi @chaoyi-wu, I notice that in the paper the max context length is set as 2048. But when I load the model, it claims that the max length is 512?
Do you have a version allowing longer sequence input? Or this one is already trained for long input, so I can just change "model_max_length" in tokenizer configurations?

[chaoyi-wu]

It is pre-trained with 2048, but instruction tuned with 512. It will be totally ok to expand the "model_max_length" if you are intended to fine-tune our mode while if you are going to perform zero-shot prompting, it can also work but may hurt the model performance.