is Alibi version available for fine tuning to a large context window?

#46
by run - opened

Current max_model_input is 2048, I need to extend it for my user case, thought alibi version of the model would be of helpful, thanks!

I noticed this as well. The code supports Alibi, but it seems that they did not use it for training this model. I wonder why. Are there downsides of Alibi that I might not know?

How do we Tun inference on the model with Alibi?

If the weights are not released that are trained with Alibi I do not belive you can. Perhaps with significantly more pretraining, it could work to some capacity.

Sign up or log in to comment