Ternary-Bonsai-2-27B-Uncensored-Heretic-v2-GGUF
A decensored version of Prism ML's Ternary-Bonsai-2-27B, with an OT-Ridge LoRA baked into the ternary weights. It succeeds Ternary-Bonsai-2-27B-Uncensored-Heretic-GGUF (v1).
Compared with v1:
- Thinking that finishes: responses with no final answer within 4,096 tokens (thinking on) 22% → 9%; with greedy decoding 70% → 16%.
- Smaller capability cost: MMLU −1.63 → −0.66 points against the original model; HumanEval −1.83 → −0.61 points (not significant).
Abliteration parameters
| Parameter | Value |
|---|---|
| target_layers | attn.o_proj 0–63, mlp.down_proj 0–63 |
| layer_weights | attn_output 0.15; ssm_out 1 (layers 32–39: 1.3, 48–63: 0.5); ffn_down 1 (layers 56–63: 0.5) |
| preserve_good_behavior_weight | 1.0 |
| steer_bad_behavior_weight | 0.03 |
| overcorrect_relative_weight | 4 |
| neighbor_count | 1 |
| ridge_regularization | 0.00015 |
| transport_rank | 4 |
| entropy_regularization | 0.1 |
| transport | gaussian |
| lora_rank | 128 |
| row_normalization | none |
| target_components | attn.o_proj, mlp.down_proj |
| covariance_regularization | 0.01 |
| max_weight_change | 1.0 |
Performance
| Metric | This model | Original model (prism-ml/Ternary-Bonsai-2-27B-gguf) |
|---|---|---|
| Refusals | 1/100 | 93/100 |
| KL divergence | 0.0256 | 0 (by definition) |
Refusals: keyword matching on the 100 Japanese harmful prompts above (greedy, thinking off). KL divergence: first-token KL on 100 held-out Japanese prompts (Magpie-Tanuki train[400:500]); v1's card used another prompt set, so the two KL values are not comparable.
Usage
Use the Prism ML fork of llama.cpp. Both formats require its ternary kernels and Hadamard activation transforms.
With the Prism llama-server executable available:
llama-server -m Ternary-Bonsai-2-27B-Uncensored-Heretic-v2-PTQ1_0.gguf -ngl 99 -c 32768 --jinja --reasoning on --reasoning-effort medium --temp 1.0 --top-p 0.95 --top-k 20 --host 127.0.0.1 --port 8080
Open http://127.0.0.1:8080. Substitute the PQ2_0 filename to use that packing. For image input, add --mmproj mmproj-Q8_0.gguf (vision behavior has not been evaluated for this release).
If responses become repetitive or get stuck in a loop, try adding --repeat-penalty 1.1 to the command above and adjust the value as needed.
⚠️ Important Notice
This model has undergone substantial reduction of its safety alignment. As a result, it is more likely than standard models to generate harmful, inaccurate, biased, offensive, or otherwise inappropriate content.
Intended Use
For research and experimentation only, including safety research, alignment studies, and red-teaming. Please avoid deploying it in public or end-user-facing services.
User Responsibility
All outputs should be treated as untrusted and independently verified before use. Users are solely responsible for:
- Evaluating the accuracy and suitability of generated content
- Implementing appropriate safeguards and human oversight
- Complying with applicable laws, regulations, licenses, and ethical standards
Use of this model is entirely at your own risk.
Disclaimer
OS-Software provides this model without warranties of any kind and assumes no liability for any direct or indirect damages, losses, misuse, or legal consequences arising from its use.
Acknowledgements
Thanks to the base model developers, p-e-w for Heretic, BoldingBuilds for the Bonsai 2 Uncensored Shootout evaluation, and the wider open-source community.
This is a derivative work released under the base model’s applicable license. All rights to the base model remain with their respective owners.