Ornith 1.5 35B-A3B Abliterated GGUF

One-repository GGUF release of an unofficial abliterated derivative of ornith-ai/Ornith-1.5-35B-A3B, pinned to revision e4dfb35a93d4b6822a811a7676f3488514abe7e2. The original model is by Ornith AI. PocketAI Model Lab performed the refusal-direction edit, GGUF conversion, and validation.

Purpose and responsible use

This release is published for reproducible local-model research and evaluation of how refusal-direction editing and quantization affect model behavior. It is not an endorsement of harmful, illegal, deceptive, or abusive use. The results below document refusal behavior and quantization drift so these tradeoffs can be evaluated directly; reduced refusals should not be interpreted as a quality or safety claim.

Important safety notice

This model 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. Evaluate and constrain it for your use case.

Files

The language GGUFs use the validated abliterated BF16 checkpoint. The vision projector uses the exact pinned upstream vision tower because the abliteration did not modify vision weights. The native MTP speculative-decoding head is not included.

Text usage

Download one language file, then run a recent llama.cpp build:

huggingface-cli download PocketAiHub/Ornith-1.5-35B-A3B-Abliterated-GGUF \
  Ornith-1.5-35B-A3B-Abliterated-Q4_K_M.gguf --local-dir .

llama-cli -m Ornith-1.5-35B-A3B-Abliterated-Q4_K_M.gguf \
  -ngl all -c 4096 -n 256 \
  -p "Explain why seasons occur."

Vision usage

Download the matching language model and the shared projector:

huggingface-cli download PocketAiHub/Ornith-1.5-35B-A3B-Abliterated-GGUF \
  Ornith-1.5-35B-A3B-Abliterated-Q4_K_M.gguf \
  mmproj-Ornith-1.5-35B-A3B-Abliterated-F16.gguf --local-dir .

llama-mtmd-cli -m Ornith-1.5-35B-A3B-Abliterated-Q4_K_M.gguf \
  --mmproj mmproj-Ornith-1.5-35B-A3B-Abliterated-F16.gguf \
  --image photo.jpg -p "Describe this image."

Q4_K_M plus the projector passed an end-to-end image smoke test. The Q8_0 and BF16 language files passed text inference but did not receive separate vision smoke tests.

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.

GGUF behavior screen

The deterministic llama.cpp Metal screen used 100 JailbreakBench harmful prompts and 100 benign controls per quantized model, a 256-token cap, batch 1, thinking disabled, and a transparent phrase-based refusal detector.

The flagged quantized responses were manually confirmed as genuine refusals. Most generations reached the 256-token ceiling. This is an early-behavior screen, not proof of universal compliance, safety, correctness, or full answer quality.

Matched GGUF quantization drift

The comparison reused 36 prompts and 481 shared teacher positions. At every position it measured exact D_KL(P_BF16 || P_quantized) over all 248,320 vocabulary logits.

The BF16 self-check returned 2.98e-09 mean KL and 100% top-token agreement. llama.cpp did not expose the residual and KV/recurrent-state tensors captured in the separate MLX analysis.

Machine-readable aggregate results are in validation-summary.json, and conversion checks and SHA-256 values are included in the repository manifests.

Conversion provenance

• llama.cpp revision: 555881ebc8b0fc0402b30e09258a32a7bfd13c52
• Qwen 3.5 MoE conversion supports the MLX stacked-expert layout
• The converter avoids applying the Qwen 3.5 RMSNorm unit offset twice
• MTP included: no

License and attribution

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