This is a merge of
LongAlpaca-70B-lora into Xwin-LM's
Xwin-LM-70B-V0.1, replacing the embed and norm layers as described in the
LongLoRA repo, and removing the extra row and pad token so that the vocabularies match.
There is no additional fine-tuning. The resulting model seems to not be broken... you can test whether it is truly the original model + 32K capability (use linear rope scaling 8).
You could also try merging this with other models of longLORA descendency (like
Aurelian).
See
this discussion for how to create merges like these.
