lemon07r

/

Gemma-2-Ataraxy-9B

like 74

What would it take to brew a 27b version of this?
Could it be like:

1. Train gemma-2-27b-it on gemma2-ultrafeedback-armorm dataset, to get model A.
2. Train gemma-2-27b-it on Snorkel-Mistral-PairRM-DPO-Dataset dataset, to get model B.
3. Train model B on gutenberg-dpo-v0.1, to get model C.
4. Merge model A and model C.

I'm asking because, I still consider this the best open source model for creative writing, bar none.
More parameters could possibly result in benefits from 27b base knowledge, and improve its long context coherence.