This is the same as Yukang's
Llama-2-70b-longlora-32k, except that the extra pad token has been stripped from the tokenizer to make it similar to the base Llama model (and it has been merged into the base model). Please refer to that page for more details.
It was created by merging
LongAlpaca-70B-lora into
Llama-2-70b, replacing the embed and norm layers as described in the
LongLoRA repo, and removing the extra row and pad token.
This is not an instruct-tuned model, but a base model for further fine-tuning. It supports 32K of context with linear rope scaling of 8.
