Serious question: What's going on here (in general)? 😳
Hello guys, thank you for this new model. I guess it is based on Llama 3.3 8B? That's very nice and appreciated! ❤
I have a question though. Are you guys like the last the Mohicans of E/RP finetuning? 😥
Seriously, this is concerning, because up to this day, I haven't seen anyone else do any finetune of Ministral 14B except you. And that's a very good model, I thought E/RP finetuners would absolutely love its ability to follow instructions!
Also, so far I haven't seen anyone do any finetune of Llama 3.3 8B except you.
Now, I know that Llama 3.3 8B was released coughleakedcough just yesterday, but if it follows the same path as Ministral 14B, then there won't be too many finetunes for it either.
What's going on? Is E/RP finetuning dead? 😥
Yes, finetuning is dying, nowadays many models are increasingly benchmaxxed and focused on assistant/agent tasks. It's sad.
Part of it is general attention fatigue.
Speficially here we find diminishing returns of novelty to biasing a small model towards different styles of prose when no new capability is being introduced.
There's also some kind of delicate flower of emergent reasoning power that the best small frontier models manage to express, and which gets 'fuzzed' by post-training processing (merges, distils, reaps). My tests for logical inference reveal this quite clearly.
Just my $.02