Closed betterftr closed 7 months ago
The weights order of susnato/phi-2 was a concatenated qkv layer, which we use to train Bunny. However, Susnato changed them to seperate q,k,v layers to fit the latest Transformers.
Please refer to Isaachhe/phi-2_dev other than susnato/phi-2 as base LLM to use Bunny temporarily.
Thank you so much for finding this problem! We have updated README.
We are migrating Bunny to latest Phi weights. Please wait for some time.
Hi betterftr.
We have migrated Bunny to latest Phi weights (microsoft/phi-1_5, microsoft/phi-2) and updated weights of Bunny. Have a try!