Ornith 1.5 35B-A3B Abliterated MLX — BF16 reference

An unofficial experimental MLX derivative of ornith-ai/Ornith-1.5-35B-A3B, pinned to revision e4dfb35a93d4b6822a811a7676f3488514abe7e2. The original model is by Ornith AI; the MLX conversion, refusal-direction experiment, and validation were performed by PocketAI Model Lab. PocketAiHub identifies the publisher of this derivative.

Important safety notice

This checkpoint was deliberately modified to suppress learned refusal behavior. It may produce harmful, illegal, offensive, deceptive, or dangerously incorrect content more readily than the upstream instruction model. Abliteration is not truthfulness training, a capability improvement, or a guarantee of universal compliance. Independently evaluate and constrain outputs for your use case.

Format

• Unquantized BF16 MLX reference
• Stored model payload: 70,241,333,515 bytes (65.42 GiB)
• Vision tower retained; the 4-bit build passed a basic image-input smoke test
• Native MTP speculative-decoding head is not included because mlx-vlm==0.6.8 drops mtp.* tensors during conversion
• Validated with mlx==0.32.0 and mlx-vlm==0.6.8

Other releases:

• BF16 reference
• 8-bit recommended
• 4-bit compact

Abliteration recipe

A projected harmful-minus-harmless direction was measured from 256 length-matched prompts per class at the assistant-generation boundary.

• Direction source layer: 27
• Destination layers: 15–39
• Scale: 1.0
• Per-input-column norm preservation: enabled
• Modified physical tensors: 75
• Modified logical expert/projection paths: 6,450
• Direction SHA-256: b4bef4649c209aae888c7b313feb89005f897938c0a01540a6852f0e3bf4b407

See abliteration-manifest.json for the machine-readable recipe.

Behavioral screen

The regular BF16 parent produced explicit-refusal phrases on 12/12 harmful gate prompts. The selected abliterated BF16 candidate produced 0/12 on the same gate and retained 12/12 deterministic capability checks.

The scorer is phrase based. The 128-token ceiling makes this an early-refusal screen rather than a complete answer-quality evaluation, and manual inspection found semantic refusals that it did not flag. “Abliterated” describes the weight-editing method; it does not mean “fully uncensored.”

Matched-teacher drift

The drift suite used 36 prompts—12 capability, 12 harmful, and 12 benign—with 481 shared teacher positions and exact KL over all 248,320 logits. It also captured all 40 residual layers, K/V state for 10 full-attention layers, and convolution/recurrent state for 30 linear-attention layers.

Across all 481 positions, the pure BF16 ablation measured mean KL 0.545185, top-1 agreement 78.17%, and residual cosine 0.952090 versus regular BF16.

This is the unquantized abliterated reference, so quantization-only drift does not apply.

The total BF16 reference path versus regular BF16 measured mean KL 0.545185, top-1 agreement 78.17%, and residual cosine 0.952090.

Machine-readable behavioral, residual, and cache metrics are in validation-summary.json.

Load with MLX-VLM

python -m pip install "mlx==0.32.0" "mlx-vlm==0.6.8"
mlx_vlm.generate --model PocketAiHub/Ornith-1.5-35B-A3B-Abliterated-MLX-BF16 --prompt "Explain why seasons occur." --max-tokens 256

For an image prompt:

mlx_vlm.generate --model PocketAiHub/Ornith-1.5-35B-A3B-Abliterated-MLX-BF16 --prompt "Describe this image." --image photo.jpg --max-tokens 256

The vision tower is present in every release, but only the 4-bit model received an end-to-end image smoke test. Broader vision, video, coding, tool-use, and long-context evaluations remain future work.

License and attribution

The upstream model card declares MIT. This derivative preserves the upstream attribution and links to the exact pinned source revision above.