Criticism, please read
First of all the default personality: it is dry and boring, like original v3. R1 was fun schizo, r1-0528 was a bit worse, a bit more obnoxious, but this is so much much worse. No fun whatsoever. And it starts every response with "Of course!" And "That's an exellent question!". Oh and it also spams me with suicide helplines if I'm upset about it like fucking Gemma(very bad model, I hate it)!
Second, the positivity bias. OG r1 was the most neutral modern model I ever used and I loved it. It could judge things more objectively than the others. R1-0528 got a bit worse, a bit more positive. I could no longer use it to judge and criticize. v3.1 is pure fucking cancer in this regard, toxically positive. Please filter your data.
Third, the style. It's just full of Geminislop. Take Kimi K2 as an example of doing it right: they took OpenAI data filtered it from slop and got one of the best styles local has to offer period.
Fourth, the thinking, or better said the lack of it. It no longer can reason through complex problems because well, it does not reason at all! It does not reason when it thinks the problem is easy while it is not. Hybrid resoner was a mistake.
My suggestions for the future:
- Filter your data! Remove all the cancerous gemini refusals, helplines, slop and positivity bias. You don't need those. You do NOT need those. Quality>quantity.
- Split reasoner and non-reasoner like before.
- Get diverse high quality data. Either take like Kimi from OpenAI o3 with heavy filtering, or try taking some from Anthropic's Claude.
- Bonus: bring back fun and smart old R1 schizo, I liked it!
This model is a big miss. The hybrid reasoning system is trash.
It reasons very less even for complex problems and produces shit outputs. If a problem does not appear complex at first glance, it won't reason much even if the problem requires deep reasoning.