Censorship is back, shame!
#5
by
notafraud
- opened
My initial expectation of reasoning model was that it would be similar to the latest Magistral - unfortunately, 14b Reasoning is censored heavier than 14b instruct and 8b models. In general, 14b is more prone to refusals than 8b models in this series, and so far I didn't see any positive trade-off from it.
Again and again, you should stop wasting your time on refusal instructions and focus on improving the quality of your models in domains where it can compete. Instead, you are making your new models potentially less popular among people who already use and like them, while not really competing with other big players in the industry. Partnership with BFB was pathetic, given their reputation.
Shame.