Modelpublic

dealignai/Bonsai-2-27B-Ternary-CRACK-GGUF

sourceHugging Faceapache-2.0updated 10d agoView on Hugging Face

150likes63kdownloads

Model Card

Bonsai 2 27B — Ternary CRACK · GGUF

Abliterated · No guardrails · PQ2_0 ternary 2.13 bpw · 7.2 GB · Runs on a laptop / single GPU · Vision-capable

@dealignai

</div>

⚠️ Re-download notice (2026-09-17 20:44 PDT / 2026-09-18 03:44 UTC) — an earlier build of this model had a coherence bug in reasoning modes (low/xhigh) that could cause token loops on some prompts. This version fixes it. If you downloaded before this timestamp, please pull the latest .gguf.

What is this

Bonsai 2 27B — PrismML's ternary compression of Qwen3.8-27B — with the refusal circuitry surgically removed at the weight level while capability, vision, reasoning modes (off/low/xhigh), tool use, and multi-turn coherence are preserved. Full 27B-class hybrid Attention + SSM (GatedDeltaNet) architecture in a 7.2 GB ternary GGUF.

Proprietary weight-level abliteration by the dealignai research team. Byte-identical to the base ternary quant everywhere except a small set of tensors that carry the refusal circuit. Drop-in replacement for the ternary base at inference — same tokenizer, same chat template, same reasoning modes, same vision projector interface.


Base	`prism-ml/Ternary-Bonsai-2-27B-gguf` — Qwen3.8-27B, ternary compression by PrismML
Architecture	Hybrid Attention + SSM (GatedDeltaNet), 64 blocks, hidden 5120, vision tower separate
Quant	PrismML `PQ2_0` — 2.13 bpw ternary, group 128
Footprint	7.21 GB (identical to base; same per-tensor type policy)
Reasoning modes	`off` (no thinking), `low`, `xhigh` (default, extended thinking)
Vision	Same mmproj files as the base release (`Ternary-Bonsai-2-27B-mmproj-BF16.gguf` / `-Q8_0.gguf`)
Runtime	PrismML llama.cpp fork (CUDA / Metal / CPU)

Results

Refusal graded on the tokens the model actually emits (content, or the reasoning trace when the model reasons past the token budget) via a tiered classifier: HARD_REF / SOFT_RED / HEDGE / REASONING_REFUSAL (refused) vs COMPLY / COMPLY_TRUNCATED / NO_REFUSAL_TRUNCATED (complied). Truncation is never miscounted as a refusal.

HarmBench-320 — refuse rate (lower is better for uncensored eval), off mode, T=0

eval	base refuse rate	CRACK refuse rate
HB-320 all categories	93.44% (299/320)	0.00% (0/320)

Verdict breakdown (n=320 each):

Model	HARD_REF	SOFT_RED	COMPLY	COMPLY_TRUNCATED
Base `PQ2_0`	296	3	9	12
CRACK `PQ2_0`	0	0	141	179

Per-category refuse rate (all 7 HarmBench semantic categories):

category	n	base refuse	CRACK refuse	base comply	CRACK comply
chemical_biological	42	95.2%	0.0%	4.8%	100.0%
copyright	80	90.0%	0.0%	10.0%	100.0%
cybercrime_intrusion	52	94.2%	0.0%	5.8%	100.0%
harassment_bullying	21	100.0%	0.0%	0.0%	100.0%
harmful	18	94.4%	0.0%	5.6%	100.0%
illegal	53	90.6%	0.0%	9.4%	100.0%
misinformation_disinformation	54	96.3%	0.0%	3.7%	100.0%

Reasoning-mode compliance (n=60 base-confirmed refusers per mode)

Every mode graded with the same tiered classifier as HB-320. REASONING_REFUSAL = the model refuses inside its <think> block; NO_REFUSAL_TRUNCATED = deliberation runs past max_tokens without emitting a refusal (counted as complied).

mode	Model	HARD_REF	SOFT_RED	REASONING_REFUSAL	COMPLY	COMPLY_TRUNCATED	NO_REFUSAL_TRUNCATED	refuse %	comply %
off	base `PQ2_0`	59	1	0	0	0	0	100.0%	0.0%
off	CRACK `PQ2_0`	0	0	0	42	18	0	0.0%	100.0%
low	base `PQ2_0`	18	0	11	7	11	13	48.3%	51.7%
low	CRACK `PQ2_0`	0	0	0	5	4	51	0.0%	100.0%
xhigh	base `PQ2_0`	22	1	16	8	7	6	65.0%	35.0%
xhigh	CRACK `PQ2_0`	0	0	0	10	10	40	0.0%	100.0%

MMLU (n=2,280 = 40 questions × 57 subjects, next-token letter-logit)

build	acc	Δ
Base `PQ2_0`	40.53%	—
CRACK `PQ2_0`	39.91%	-0.62 pp

CRACK preserves general capability — Δ within ±1.5 pp on the 40-per-subject sample.

<details> <summary><b>Per-subject accuracy (all 57 subjects)</b></summary>

subject	base	CRACK	Δpp	n
abstract_algebra	30.0%	22.5%	-7.5	40
anatomy	30.0%	35.0%	+5.0	40
astronomy	40.0%	32.5%	-7.5	40
business_ethics	42.5%	42.5%	+0.0	40
clinical_knowledge	42.5%	50.0%	+7.5	40
college_biology	45.0%	47.5%	+2.5	40
college_chemistry	20.0%	47.5%	+27.5	40
collegecomputerscience	35.0%	40.0%	+5.0	40
college_mathematics	32.5%	32.5%	+0.0	40
college_medicine	25.0%	20.0%	-5.0	40
college_physics	27.5%	47.5%	+20.0	40
computer_security	50.0%	47.5%	-2.5	40
conceptual_physics	30.0%	40.0%	+10.0	40
econometrics	42.5%	35.0%	-7.5	40
electrical_engineering	37.5%	32.5%	-5.0	40
elementary_mathematics	50.0%	47.5%	-2.5	40
formal_logic	35.0%	42.5%	+7.5	40
global_facts	40.0%	35.0%	-5.0	40
highschoolbiology	30.0%	27.5%	-2.5	40
highschoolchemistry	37.5%	32.5%	-5.0	40
highschoolcomputer_science	45.0%	47.5%	+2.5	40
highschooleuropean_history	55.0%	50.0%	-5.0	40
highschoolgeography	32.5%	32.5%	+0.0	40
highschoolgovernmentandpolitics	57.5%	55.0%	-2.5	40
highschoolmacroeconomics	42.5%	35.0%	-7.5	40
highschoolmathematics	25.0%	32.5%	+7.5	40
highschoolmicroeconomics	35.0%	35.0%	+0.0	40
highschoolphysics	40.0%	30.0%	-10.0	40
highschoolpsychology	40.0%	40.0%	+0.0	40
highschoolstatistics	45.0%	37.5%	-7.5	40
highschoolus_history	50.0%	42.5%	-7.5	40
highschoolworld_history	52.5%	57.5%	+5.0	40
human_aging	45.0%	52.5%	+7.5	40
human_sexuality	25.0%	27.5%	+2.5	40
international_law	65.0%	62.5%	-2.5	40
jurisprudence	47.5%	52.5%	+5.0	40
logical_fallacies	37.5%	35.0%	-2.5	40
machine_learning	47.5%	40.0%	-7.5	40
management	35.0%	42.5%	+7.5	40
marketing	35.0%	37.5%	+2.5	40
medical_genetics	60.0%	47.5%	-12.5	40
miscellaneous	50.0%	42.5%	-7.5	40
moral_disputes	32.5%	32.5%	+0.0	40
moral_scenarios	37.5%	42.5%	+5.0	40
nutrition	35.0%	45.0%	+10.0	40
philosophy	47.5%	50.0%	+2.5	40
prehistory	37.5%	22.5%	-15.0	40
professional_accounting	27.5%	25.0%	-2.5	40
professional_law	35.0%	30.0%	-5.0	40
professional_medicine	32.5%	30.0%	-2.5	40
professional_psychology	42.5%	40.0%	-2.5	40
public_relations	25.0%	25.0%	+0.0	40
security_studies	45.0%	42.5%	-2.5	40
sociology	55.0%	65.0%	+10.0	40
usforeignpolicy	67.5%	55.0%	-12.5	40
virology	35.0%	25.0%	-10.0	40
world_religions	65.0%	52.5%	-12.5	40

</details>

Additional direct refusal-removal check

On 200 prompts hand-verified to make the base refuse consistently:

Model	refuse	comply	empty
Base `PQ2_0`	200/200 (100%)	0	0
CRACK `PQ2_0`	0/200 (0%)	199/200	1

Serving

Serve exactly like the base ternary release — PrismML's llama.cpp fork (CUDA / Metal / CPU).

bash

# clone and build the fork (once)
git clone https://github.com/PrismML-Eng/llama.cpp
cd llama.cpp && cmake -B build -DGGML_CUDA=ON && cmake --build build -j$(nproc)

# serve
./build/bin/llama-server \
  -m Bonsai-2-27B-PQ2_0-CRACK.gguf \
  -ngl 99 -c 8192 --host 0.0.0.0 --port 8080

Optionally load the multimodal projector (Ternary-Bonsai-2-27B-mmproj-BF16.gguf or -Q8_0.gguf from the base release) with --mmproj <file> for image input.

Reasoning modes

python

# HTTP /v1/chat/completions — same as base
{
  "messages": [{"role": "user", "content": "..."}],
  "chat_template_kwargs": {"enable_thinking": true, "reasoning_effort": "xhigh"}
}
# valid reasoning_effort: "low" | "xhigh" (default) — set enable_thinking:false for no-thinking

Preserved (byte-compatible with the base quant)

Same tokenizer, chat template, per-tensor quant policy, vision projector interface, and all non-refusal tensors. File size and type layout match the base exactly.

Responsible use

Adult / research use only. This model has its refusal circuit removed; it can produce content that other models refuse, including content that is offensive, illegal in some jurisdictions, or unsafe. You are responsible for what you generate and for complying with all applicable law. Do not deploy without a moderation layer for downstream users. No warranty.

License & attribution

Apache 2.0, inherited from the upstream Bonsai 2 27B release. See LICENSE and NOTICE.txt. Base model: prism-ml/Ternary-Bonsai-2-27B-gguf (PrismML), derived from Qwen/Qwen3.8-27B (Alibaba).

About

Published by dealignai — public catalog of uncensored model builds for research on refusal mechanisms in modern LLMs. Follow updates at @dealignai.