[Request] Train a Medusa version of the model using the original training data
#58
by
narai
- opened
Medusa is the best 7B open source model, but it is possible to double the inference speed. This is easiest done with the original training data that the Mistral team has.
Alternatively, I might try to bootstrap the model by using Mistral to create my own training data for my use cases. This won't work as well as what could be achieved with the highly diverse training data of the original train set.