Qwen3 32B
Are you considering https://huggingface.co/nvidia/Qwen3-Nemotron-32B-RLBFF or just the base qwen 3 32B as a good base to train on? The nemotron seems decent enough at RP/storywriting out of the box (thinking turned off) so I'd be really interested to see what it could do trained on your dataset.
Hey,
No, hadn't seen this version of the Qwen3 32b. I gave it a quick train while I had my stuff setup, but not really liking the result. One of the problems with Qwen3 initially from memory was the censorship and I think nvidia maintained that pretty well unfortunately.
The GGUF is here though if you'd like to test for yourself, maybe for your creative tasks it works better: https://huggingface.co/ApocalypseParty/Qwen-32B-Nemotron-SFT-2-Q6_K-GGUF
Whoa, thanks for at least trying! Shame it won't work out but I'll at least spin it up and see what it can do. Thanks for taking the time!
One thing maybe worth investigating is the Abliterated-version of the models. I found decent luck with Qwen 3 once abliterated to be quite uncensored.
On another note, when you say that "Nvidia maintained that pretty well" what do you mean by that?
Qwen is a pretty censored model in my experience. Nvidia have quite a lot of safety data in their datasets too, so them training on top of Qwen didn't do anything to reduce the existing censorship is what I meant. I could try an abliterated model, but personally I'd rather just use a mistral upscale or GLM 4 32b instead.