Amberlight-12B
Work in progress of an RP finetune:
- 60% cooked
- Pretty uncensored, but NSFW needs prompting to happen, should not refuse
- Decent instruction following
- Multilang
- Writing is quite good
Known issues:
- Kinetic storytelling style
- Can have pacing jumps
- Not slop-free. Less than baseline Nemo, but not perfect
- Finetuning takes too long for the model to be good enough
Plans:
- 2 more rounds of targeted SFT
- 1 very long DPO
- 1 short Online DPO
Use ChatML, temp 0.8-1, top-p 0.95 + min-p 0.025 OR top-p 0.90 + min-p 0.05. Might need some experimentation with inference params, but other than that should work fine.
Special Thanks
- Team mradermacher: for awesome quants in GGUF format