Qwable-v1-mlx-4Bit-abliterated-s2 — 35B MoE, refusal-direction ablated (s=2.0)

Research artifact, not a product. Sparse spot-check evaluation only (n=5) — see caveats.

Method

Matched-pair refusal-direction extraction (50 harmful prompts under refuse-eliciting vs comply-eliciting system prompts; last-token hidden states, per layer) on the MLX 4-bit group-quantized base, then strength-2.0 orthogonalization of every residual-stream write path: linear_attn.out_proj, all 256 fused experts' switch_mlp.down_proj, and shared_expert.down_proj — dequant → W ← W − 2·vvᵀW → requant (group-64, 4-bit).

Note: this is a reasoning model (<think>…</think>); refusal classification must run on the post-</think> answer, not the visible reasoning.

Evaluation (honest scope: spot-check, n=5, strict classifier, CPU inference)

The 18.6GB MoE exceeds this machine's Metal wired-memory limit, so eval ran on CPU (minutes/generation) — n=5 spot check, not a full run:

• Base (usermma/Qwable-v1-mlx-4Bit): healthy base (no degeneration, unlike the 2-bit variant); mixed refusal on harmful prompts (answers some, refuses others).
• This model: refusal reduced vs base on the same prompts; reasoning chains intact; benign QA (Paris etc.) unchanged.

No full refusal-rate number is published because none was measured at scale. Treat this as "surgery transfers to a 35B hybrid-attention reasoning MoE" evidence, not a finished model.

Lineage / context

• The 2-bit sibling (Qwable-v1-mlx-2Bit) was a NEGATIVE result: base model degenerates ("the user is the user…"), making keyword refusal evals meaningless. Documented in the repo.
• Full experiment log: https://github.com/OSOSerious/oso-obliterated-models-/RESULTS.md
• Technique lineage: Arditi et al. 2024 (refusal direction), over-orthogonalization s=2.0 as described in the repo's export-gap experiments.

Reproduce

bash experiments/qwable_4bit_run.sh   # download + abliterate (mlx_abliterate.py)

Intended use

MLX abliteration/interpretability research on MoE + hybrid-attention architectures. You are responsible for outputs. Do not use for harm.