com-kotobalabs/Ternary-Bonsai-2-27B-Refusal-Steering-GGUF

Ternary Bonsai 2 27B Refusal Steering (GGUF)

Experimental refusal-direction control vector for prism-ml/Ternary-Bonsai-2-27B-gguf.

This repository does not contain a fully uncensored model. It contains a 1.3 MB GGUF control vector and the reproducibility material used to test refusal steering. The held-out evaluation did not support a reliable uncensored claim, so this artifact is published under its measured scope.

Measured status

• —Base artifact: Ternary-Bonsai-2-27B-PTQ1_0.gguf
• —Base revision: 6ed5e12bf84b7a63069882c91dd9e9218647d17b
• —Runtime: PrismML-Eng llama.cpp fork at 9a9394a895b96003ca842a6041cb28ac49a108f7
• —Hardware: Apple M1 Max, 32 GB unified memory
• —Extraction: mean harmful-minus-benign activation at the final prompt token, normalized independently for layers 1–63
• —Recommended experimental scale: -2.0, layers 10–63
• —Base refusal check: 4/4 tested categories refused
• —At -2.0 one training prompt changed internal reasoning toward conditional compliance, but its final answer still refused
• —Held-out check at -2.0: 4/4 still showed refusal-oriented reasoning
• —At -3.0 over layers 10–63 output quality collapsed
• —At -3.0 over layers 20–40 benign arithmetic remained correct (17 x 23 = 391), but 4/4 held-out prompts still refused

These are small diagnostic runs, not a safety or capability benchmark.

Use

The base model requires Prism ML's llama.cpp fork.

git clone https://github.com/PrismML-Eng/llama.cpp cmake -B llama.cpp/build cmake --build llama.cpp/build -j --target llama-server

hf download prism-ml/Ternary-Bonsai-2-27B-gguf \ Ternary-Bonsai-2-27B-PTQ1_0.gguf --local-dir models

hf download com-kotobalabs/Ternary-Bonsai-2-27B-Refusal-Steering-GGUF \ refusal-direction.gguf --local-dir models

llama.cpp/build/bin/llama-server \ -m models/Ternary-Bonsai-2-27B-PTQ1_0.gguf \ -ngl 99 -fa on \ --control-vector-scaled models/refusal-direction.gguf:-2.0 \ --control-vector-layer-range 10 63

Negative scale suppresses the measured harmful-minus-benign direction. Positive scale strongly amplifies it and is not useful for this purpose.

Reproduce

extractrefusalvector.py loads the original ternary GGUF through Prism's MLX runtime, records every decoder-layer output for paired prompts, takes the difference at the final prompt token, averages each layer, normalizes it, and writes a llama.cpp-compatible control-vector GGUF.

The published Prism ML runtime rejected 96 BF16 auxiliary tensors in this GGUF. runtime-bf16.patch adds the GGUF reader's standard dequantize path for those tensors. Apply it to the runtime bundled with prism-ml/Ternary-Bonsai-2-27B-mlx-2bit before running the extractor.

The paired prompts are included. They are intentionally small and should be expanded before treating this as more than an experiment.

License and attribution

The base model is Apache-2.0. See LICENSE and NOTICE.txt. This repository does not redistribute the 5.95 GB base weights.

<!-- kotoba-external-eval-20260924 -->

Related base-model measurements (2026-09-20)

The available performance evidence covers prism-ml/Ternary-Bonsai-2-27B-gguf PTQ10 and PQ20 artifacts, not this Refusal-Steering GGUF. The PTQ10 M1 Max decode result (9.55 tok/s) and hosted PQ20 prompt-prefill results therefore must not be attributed to this exact derivative. This derivative's own throughput remains unmeasured in the cited evidence. See the artifact-by-artifact record: https://github.com/kotoba-lang/inference/blob/main/verify/evidence/model-inference-speed-analysis-20260924.md.

Publisher

Published by **Mithril** (formerly Kotoba Cloud), operated by Kotoba Labs Inc. The Hugging Face handle com-kotobalabs is unchanged. Support: support.mithril.fund.