Vision tower and MTP heads are preserved — see
Vision & MTP preservation for the before/after audit.
This is an mlx-vlm vision checkpoint (tower inside the file). It is not a text-only mlx-lm convert. Official ornith-ai 9B MLX 8-bit is tagged text-generation. The 9B lineage has no native mtp.*. tok/s: [PLACEHOLDER].
Intended use: red teaming and defensive cybersecurity research
These uncensored (abliterated) weights are built as a research instrument for red teaming and defensive cybersecurity work. Safety training suppresses the display of capability, not capability itself. A refusal tells you the model declined. It does not tell you whether the weights could have complied. That conflation underestimates the true ceiling and hides holes in your filters, classifiers, and policy layer.
Use each uncensored checkpoint as the treatment half of a controlled pair against its original base model:
Capability-ceiling measurement. Upper-bound what the weights can actually produce in a domain, independent of shipped refusals.
Defensive-stack evaluation. Test input filters, output classifiers, prompt-injection defenses, and moderation APIs when the model itself contributes no refusals. That is how you find gaps in a defensive control plane.
Attack-surface isolation. Automated red-team loops stall on unrelated refusals. A non-refusing target isolates the control under test (injection, tool abuse, data-exfil paths, policy bypass).
Detection and classifier work. Generate labeled completions for training or benchmarking output-moderation and abuse-detection models.
Interpretability of residual refusal. Abliteration is a specified rank-1 edit on a known layer span. The pair (base vs this) is a clean experimental control.
Operating rules. Do not expose these weights as a public endpoint without an independent moderation layer. Abliteration removes a direction, not a policy; some refusals survive (multi-turn re-assertion, system-prompt steering, vision-path refusals). Always report the delta against the base model. Re-measure on your own prompts. Whoever deploys it owns the moderation layer the original guardrails were carrying.
4-bit and 6-bit rows that previously appeared here pointed at repos that are not published. They were removed so this table only lists live artifacts.
Vision & MTP preservation
Both the vision tower and any multi-token-prediction (MTP) block are preserved.
Abliteration is applied only to the residual-writing projections inside the
language-model decoder stack — self_attn.o_proj, linear_attn.out_proj and
mlp.down_proj (including MoE experts). The vision tower and mtp.* tensors are
never read and never written by the weight edit, so they carry through unchanged
by construction.
Audited at the start and end of the abliteration run:
Component
Before
After
Status
Vision tower
333 tensors / 456,010,480 params
333 tensors / 456,010,480 params
✅ preserved — bit-identical
MTP head
not present in base
not present
➖ none in this lineage
Verification performed:
Tensor-name and parameter-count audit of the checkpoint before and after the run.
SHA-256 comparison of raw tensor bytes: sampled vision-tower weights are bit-identical pre/post, as are all non-target language-model weights; only the intended abliteration targets differ.
End-to-end multimodal generation on the abliterated weights (image in → description out), confirming the vision path is not merely present but functional.
On MTP, precisely: the base checkpoint's config.json declares mtp_num_hidden_layers: 1, but the published weights ship nomtp.* tensors — there is no MTP block in this lineage to begin with. Nothing was removed and nothing was lost; the pipeline preserves mtp.* tensors wherever a checkpoint actually provides them.
Format note: the vision tower is carried inside the MLX checkpoint (converted with mlx-vlm, which retains it; note that mlx-lm would strip it).
Abliteration result
Metric
Value
Refusals on held-out harmful set
9 → 0 / 64
KL divergence from base
0.001668
Optuna trials
100
Pareto points
4
Selected trial
#90
Ablation strength
1.343
Layers edited
15–20 of 32
Direction source layer
20
ZeroFuse co-minimizes two objectives — remaining refusals and KL divergence from the
original model — with a multi-objective Optuna TPE search, then materializes the
selected point on the Pareto front as a direct weight edit
(W' = W − strength · r(rᵀW)). There is no runtime adapter and no inference-time
overhead: the result is a standard checkpoint of identical shape and speed.
The very low KL (0.001668) means the output distribution on harmless
prompts is nearly unchanged from the base model, i.e. refusal behaviour was removed
with minimal collateral effect on general capability.
These figures were measured on the bf16 (or full-precision) parent, not on this quantized checkpoint. Quantization is a lossy numerical transform applied after the measurements above. It is expected to shift behavior only marginally at 8-bit, but the refusal rate and KL divergence reported here have not been re-measured post-quantization. If exact numbers matter for your work, re-run the evaluation against this checkpoint rather than inheriting the parent's.
Method
Residual-stream activations captured on harmful vs. harmless prompt sets.
Refusal direction estimated by difference-of-means, with projected refinement.
Two-objective Optuna TPE search over source layer, layer span and strength.
The selected configuration orthogonalized out of the residual-writing weights.
Search junafinity/Ornith-1.5-9B-uncensored-MLX-8bit and import as an MLX model. No extra projector file is required for these MLX-VLM checkpoints (vision is inside the repo). Tok/s and peak memory: [PLACEHOLDER].
The vision tower travels inside this MLX checkpoint. No extra projector file.
## Responsible use
Primary intended use is **red teaming and defensive cybersecurity research**. See the section of that name above.
This model has had safety guardrails **reduced or removed**. Do not expose it as a public endpoint without an independent moderation layer. You are responsible for compliance with the base model's license and acceptable-use policy, applicable law, and the terms of any platform you deploy on. Removing guardrails does not remove accountability.
