stickmouse-Ornith-1.0-9B-abliterated-Fable-Opus-4.7-Distilled-SFT

Stickmouse-Ornith-1.0-9B-abliterated-Fable-Opus-4.7-Distilled-SFT

First-ever finetune done entirely on local hardware. This is Ornith-9B finetuned on a Claude Opus/Fable-distilled dataset — as far as I know, no one else has finetuned Ornith-9B on this kind of dataset before.

Training details

• Base model: YuYu1015/YuYu1015-Ornith-1.0-9B-abliterated — thanks to YuYu1015 for the abliteration, saved me from having to do it myself.
• Method: QLoRA
• Steps: 3546 steps, targeting 2–3 epochs
• Hardware: Tesla P100 16GB + Titan V 12GB (2016–2018 era cards)
• Status: Training was interrupted by storage running out at ~71% completion. The released checkpoint reflects that partial run, but still performs well.

This was built on a budget with old hardware — no cloud compute, no big GPU rental. If you're working with even more limited hardware and know how to make it work, I'd love to hear from you.

Planned future finetunes

• Qwen3.5-9B
• Attempting Qwen3.6-35B-A3B via QLoRA (ambitious given the hardware, but going to try)

Quantized versions

GGUF quants (f16, Q8_0, Q4_K_M) are available at: Stickmouse/stickmouse-Ornith-1.0-9B-abliterated-Fable-Opus-4.7-Distilled-SFT-GGUF