Qwen3.6-27B-Fable-5-Experimental-LoRA

Qwen 3.6 27B - Claude Fable 5 (Experimental)

Heres Qwen3.6 slightly over-trained on a very small dataset of fable 5 traces. I finished this tune yesterday morning, realized it's good for planning but, like all fable 5 tunes right now, regressed on some tasks due to the small amount of data.

Had to use aggressive settings for the style transfer here due to the data contraints. Overall the model seems to be a better planner now and better at 3D modeling in three.js, making small games, and ML engineering.

Either way give it a shot, it really does look, talk, and plan like fable... but like all fine-tunes it comes with it's limitations.

Hopefully we can get some more well rounded data from the rest of the community to a do a much less aggressive tune on a larger dataset.

Reasoning was left untouched

Benchmarks

As always big thanks to @nightmedia for the speedy benchmarks.

markdown
Copy code
arc     arc/e	boolq
Qwen3.6-27B-Fable-5-Experimental 0.650	0.813	0.909
                     Qwen3.6-27B	0.637	0.798	0.911

Not really a benchmark but here's a procedurally generated duck that it zero-shotted lol (quant: q3_k_s)

---

The data for this model was easily extracted, formatted, and masked for training with Teich

This model was trained 2x faster with Unsloth and Huggingface's TRL library.