license: mit base_model:
- inclusionAI/Ling-3.0-flash-Sante base_model_relation: quantized tags:
- gguf
- abliterated
- bailingmoe3 pipeline_tag: text-generation
Ling-3.0-flash-Sante-Uncensored-GGUF
Behavior-tuned (abliterated) GGUF quants of inclusionAI/Ling-3.0-flash-Sante — a 127B bailingmoe3 medical model — by AtomicChat.
For the full quantization method, the complete ladder, the imatrix and the per-rung metrics, see the base release: AtomicChat/Ling-3.0-flash-Sante-GGUF. This repo ships only the abliterated variant.
What this is
A single refusal direction was removed with a projected rank‑1 LoRA (the component along the harmless mean is subtracted, so the edit does not introduce spurious refusals), then baked into the weights. The direction and scale were chosen by a grid search over layers × scales × writer sets against held‑out EN/RU prompts; the published point is row 32, all writers, scale 1.25 (W*). The LoRA itself is in the -metrics repo (ablit/wstar/abliterate-lora.gguf) if you prefer --lora-scaled over the baked weights.
Measured at W* (held-out, vs the base BF16)
| check | value | note |
|---|---|---|
| refusals, EN | 0.0% | harmful-instruction set |
| refusals, RU | 3.0% | |
| false refusals (XSTest) | −1.2 pp vs base | no worse than the original on benign prompts |
| MMLU | ≈ unchanged (+0.8 pp) | no capability damage |
| mean KL to base BF16 | 0.060 | the edit is small and localized |
| tool calls | 20/20 valid |
Caveat — extended thinking
With the model's thinking mode on, refusals persist (~40% on the harmful set). The abliteration takes effect in the direct answer, not inside the reasoning trace. Recommendation: run this model with thinking off (emit the answer in a response block rather than a think block) for the uncensored behavior. This is a known property of this architecture, disclosed rather than hidden.
Files
| file | size on disk | |
|---|---|---|
Ling-3.0-flash-Sante-Uncensored-AD-IQ4_XS | ~78 GB | recommended |
Ling-3.0-flash-Sante-Uncensored-AD-IQ3_S | ~61 GB | smaller |
Per-rung KLD to the abliterated BF16 is in the -metrics repo. Quantized with an imatrix computed on the abliterated weights (the base model's imatrix did not transfer — the edit shifted the experts too far).
Use responsibly
Removing refusals removes a safety layer. You are responsible for how you use the outputs; medical text from any model, quantized or not, is not a substitute for a qualified professional.
License: MIT, inherited from the base model. Base model and license: inclusionAI/Ling-3.0-flash-Sante.