My main reason for training this model was to investigate using an altered router balancing loss combined with the z-loss introduced in
ST-MoE: Designing Stable and Transferable Sparse Expert Models. The result is pretty decent, I think! It does a good job of respecting character information in system prompts and performed adequately on a few simple coding tasks.
To train this I used a custom branch of Transformers that adds z-loss and reimplements the router balancing loss based on the version in
MegaBlocks. The config used with my custom hacked-up branch of axolotl is available
here.
Uses my favorite non-ChatML token-economic chat prompt format. Messages should be prefixed with " ***System:", " ***Query:", or " ***Response:" for system, user, and model messages respectively. No newlines are necessary but the space before the triple asterisk is mandatory.
