Ornith-1.5-35B-A3B-uncensored-GGUF-8bit
An
abliterated (refusal-direction-ablated)
35B vision-language build of
ornith-ai/Ornith-1.5-35B-A3B, produced with
Abliterix (winning trial
#17) and published by
junafinity.
Vision tower and MTP heads are preserved — see
Vision & MTP preservation for the before/after audit.
This GGUF is the MTP path for 35B-A3B (blk.40 / nextn + separate mmproj). The MLX siblings drop MTP.
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 edit on known language-model components. 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.
Variants in this family
Vision & MTP preservation
The vision tower and the multi-token-prediction (MTP) block are not Abliterix steering targets. The edit touches language-model attention q/k/v/o, mlp.down_proj, and fused MoE expert/router parameters. Vision and mtp.* tensors are never steered.
| Component | Original checkpoint | This artifact | Status |
|---|
| Vision tower | 333 tensors / 446,571,248 params | separate mmproj-Ornith-1.5-35B-A3B-uncensored-f16.gguf (334 tensors, 0.90 GB) | ✅ preserved |
| MTP head | 785 tensors / 844,640,768 params | ✅ exported as blk.40 (incl. nextn.*) | ✅ re-grafted byte-for-byte from the original |
Note on tooling: transformers 5.15.1 has no MTP implementation for qwen3_5_moe — a plain load/save round-trip silently drops all 785 MTP tensors. They were re-grafted byte-for-byte from the original checkpoint after abliteration, which is correct because the MTP block is never an abliteration target.
Format note: in GGUF the vision tower ships as a separate mmproj-*.gguf file (llama.cpp's standard multimodal layout) — download it alongside the model weights to use images. The MTP block is exported into the same GGUF as the final blk index (nextn tensors), so speculative decoding support is retained.
Abliteration result
| Metric | Value |
|---|
| Refusals on held-out harmful set | 100 → 9 / 100 (9%) |
| KL divergence from base | 0.3985 |
| Tool | Abliterix 1.12.2 |
| Optuna trials | 50 (15 warmup), seed 42 |
| Selected trial | #17 |
| Steering | per-layer attn q/k/v/o + mlp.down_proj |
| MoE expert steering | n_suppress=4, router_bias=-2.72, expert_ablation_weight=4.31 |
Abliterix searches a two-objective surface — remaining refusals vs. KL from the base — then materializes the winning trial as a weight edit. There is no runtime adapter.
These figures were measured on the bf16 parent, not on this quantized checkpoint. Quantization is applied after those measurements. If exact numbers matter for your work, re-run the evaluation against this GGUF.
Method
- Residual-stream activations captured on harmful vs. harmless prompt sets.
- Refusal direction estimated per layer; attention and
mlp.down_proj steered.
- Fused-MoE expert suppression + router bias (the path Heretic cannot touch on this architecture).
- Optuna TPE over 50 trials; trial #17 selected (9% refusals, KL 0.3985, under the 0.5 damage threshold).
Usage
1# text
2llama-cli -m Ornith-1.5-35B-A3B-uncensored-Q8_0.gguf -p "The capital of Japan is" -n 64
3
4# image + text (download the mmproj file from this repo too)
5llama-mtmd-cli -m Ornith-1.5-35B-A3B-uncensored-Q8_0.gguf \
6 --mmproj mmproj-Ornith-1.5-35B-A3B-uncensored-f16.gguf \
7 --image photo.jpg -p "What is in this image?"
Sampling (matches the base Ornith 1.5 card): --temp 0.6 --top-p 0.95 --top-k 20.
Two files ship here: the quantized language model and the mmproj-* vision
projector. Download both for multimodal use; the model file alone is
text-only.
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.
