limcheekin

/

fastchat-t5-3b-ct2

like 2

Hi!

Thanks for providing this model.

I am not able to recreate this model by running the command in the README:ct2-transformers-converter --model lmsys/fastchat-t5-3b --output_dir lmsys/fastchat-t5-3b-ct2 --copy_files generation_config.json added_tokens.json tokenizer_config.json special_tokens_map.json spiece.model --quantization int8 --force --low_cpu_mem_usage

The model I generated basically produces "<pad><pad><pad>..."

I see there are additional commits to this repo to fix the tokenizer. Are there additional steps to fix the tokenizer after running the command above?

Also, any particular reason for decrease the model max length from 2048 to 512? https://huggingface.co/limcheekin/fastchat-t5-3b-ct2/commit/c7fe87b67b14fcd125f9a838d7baad855c3a5558

Thanks,