metadata
license: cc-by-nc-4.0
Mixtraln't 4x7B
Oh boy, a new model architecture in Transformers! Time to do profane things with it.
What if instead of training a MoE from scratch, we took some pre-trained Mistral models and shoved them in a little clown car?
Uses parts from the following models:
- Q-bert/MetaMath-Cybertron-Starling
- NeverSleep/Noromaid-7b-v0.1.1
- teknium/Mistral-Trismegistus-7B
- meta-math/MetaMath-Mistral-7B
- PocketDoc/Dans-AdventurousWinds-Mk2-7b
Works and generates coherent text. The big question here is if the hack I used to populate the MoE gates works well enough to take advantage of all of the experts. Let's find out!
Prompt format: maybe alpaca??? or chatml??? life is full of mysteries