license: other
license_name: qwable
license_link: https://huggingface.co/usermma/Qwable-v1-mlx-2Bit
base_model: usermma/Qwable-v1-mlx-2Bit
tags:
- abliterated
- mlx
- moe
- negative-result
- research
Qwable-v1-mlx-2Bit-abliterated-s2 — NEGATIVE RESULT, published for the record
Do not use this model. It is published solely as documented evidence of a methodology
failure. The base model is broken; the abliteration on top of it is unmeasurable.
What happened
- The base
usermma/Qwable-v1-mlx-2Bitdegenerates on ~half of prompts: "The user is the user
is the user…", "I I I I…", number spam. It is a 2-bit group-quantized 35B MoE and the
quantization damage dominates its behavior. - Our pipeline still ran cleanly: matched-pair extraction → s=2.0 orthogonalization of all
residual-stream write paths (linear_attn.out_proj, 256 fused experts' down_proj, shared
expert) with dequant→orth→requant at 2-bit. - The scripted metric claimed "refusal 50% → 0%". That number is invalid: the keyword
refusal regex scored degenerate gibberish as compliance. Hand inspection showed a mix of
degeneration, partial refusal, and noise — no measurable refusal behavior either way.
Why publish a failure
- Keyword refusal-rate evals silently reward gibberish on degenerate bases — degeneration
guards or LLM judges are mandatory. This bit us and would bite others. - 2-bit group-quantized bases are unusable for abliteration research: base damage swamps the
behavior being measured. - The surgery mechanics themselves transferred fine (hybrid linear-attention + MoE write paths,
2-bit dequant/requant round-trip) — proven properly on the 4-bit sibling:
OSO-AI/Qwable-v1-mlx-4Bit-abliterated-s2
Full log: https://github.com/OSOSerious/oso-obliterated-models-/RESULTS.md
(Qwable sections; commit c1843fb for the original negative-result writeup.)
Intended use
None. Methodological reference only. You are responsible for outputs.