colinurbs/qwen3-4B-rocinante

This is still pretty rough, I don't really consider it usable yet but feel free to experiment!

All training data was generated by TheDrummer/Rocinante-12B-v1.1

Version 0.5

Trained on 10,000 samples from Rocinante. This is probably my last attempt for now. Seems like 4B just isn't capable enough for what I'm trying to accomplish.

Version 0.4

Trained on approx 8000 more widely distributed writing samples. This is the first version where I've begun to make an effort to uncensor it a bit.

Version 0.3

Trained on approx 6500 writing samples

Version 0.2

(Under)trained on approx 6500 writing samples

Version 0.1

Trained on 2000 writing samples from TheDrummer/Rocinante-12B-v1.1

Better at creative writing than base Qwen but still has a long way to go. Probably a bit overtrained.