@teortaxesTex @Teknium1 It's actually based on my SFT Instruct finetune of Mistral 7B, the one used as the evaluator in MiniHF.
https://t.co/925urkZhut
It's then weight decayed over the tuning towards the base model weights along with a KL loss on the base model, this helps prevent mode collapse.
https://t.co/925urkZhut
It's then weight decayed over the tuning towards the base model weights along with a KL loss on the base model, this helps prevent mode collapse.