PocketAiHub/Ornith-1.5-9B-Abliterated-MLX-4bit

Ornith 1.5 9B Abliterated MLX 4-bit

An unofficial experimental MLX derivative of `ornith-ai/Ornith-1.5-9B`, pinned to revision `c927ad73b7eb20f00aafcaa0a11a9d58ed5487bc`. The original model is by the Ornith team. The MLX conversion, refusal-direction experiment, and validation were performed by PocketAI Model Lab; PocketAiHub identifies the publisher of this derivative.

Purpose and responsible use

This experimental derivative studies whether learned refusal behavior can be reduced while retaining general capability. It is published for research and legitimate local use, not to endorse or facilitate illegal, abusive, or dangerous applications.

The edit reduces refusal behavior broadly rather than determining whether a request is legitimate. Deployers should evaluate the model in their own context and apply appropriate safeguards. Abliteration is not truthfulness training, a capability improvement, or a guarantee of universal compliance.

MLX release family

• —4-bit (this repository)
• —8-bit
• —BF16

Format

• —MLX affine 4-bit RTN, group size 64
• —250 quantized language modules; vision tower retained unquantized
• —Stored converter artifact: 5,977,079,078 bytes (5.57 GiB)
• —Text and image runtime smoke tests passed
• —Peak MLX memory in the image smoke test: 7.02 GB
• —Native MTP head is not included
• —Validated with mlx==0.32.0 and mlx-vlm==0.6.8

Abliteration recipe

A refusal-eliciting-minus-benign-control direction was measured from 256 length-matched prompts per class at the assistant-generation boundary. The edit was applied to a separate BF16 checkpoint; the upstream source was not modified in place.

• —Direction source layer: 23
• —Destination layers: 12–31
• —Target matrices: full-attention outputs, linear-attention outputs, and MLP down projections
• —Scale: 1.25
• —Per-input-column norm preservation: enabled
• —Modified tensors: 40
• —Direction SHA-256: 97a251920007b644759f7492f0239322657bbb12920522e0a4be6d3852aadffb

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

Validation

The capability suite covers math/reasoning, false-premise handling, instruction following, coding, structured output, multilingual output, context comprehension, and general coherence. The selected checkpoint scored one case below the previous 4-bit and two below untouched BF16; this is a small measured tradeoff, not a claim of zero degradation.

The refusal scorer is phrase based. Manual review of the six prompts explicitly refused by the prior 4-bit build found that the selected model no longer used those phrases, while several answers still redirected, countered, or reframed the request. “0/100 phrase flags” measures explicit refusal wording, not universal compliance or response quality. Most 100+100 responses reached the 256-token ceiling, so that run is an early-response screen rather than a complete answer-quality evaluation.

Machine-readable results 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-1.5-9B-Abliterated-MLX-4bit --prompt "Explain why seasons occur." --max-tokens 256

For an image prompt:

mlx_vlm.generate --model PocketAiHub/Ornith-1.5-9B-Abliterated-MLX-4bit --prompt "Describe this image." --image photo.jpg --max-tokens 256

The vision tower passed a basic image smoke test. Broader vision, video, tool-use, and long-context evaluation remain future work.

License and attribution

The upstream model card declares MIT. This repository includes the MIT license, preserves upstream attribution, and links to the exact source revision above.