---
language:
- en
license: cc-by-nc-sa-4.0
---

# Frankenstein-MoE

### Method
To initialize the gate projection weight of the MoE layer, the H6 trainset was sampled and used. We sampled 400 and selected the final 30 with low PPL.  

trufulqa used gpt4 to generate data.

### Evals
in progress