license: apache-2.0 base_model:

• BoldingBuilds/Ternary-Bonsai-2-27B-Abliterated-PTQ1_0-GGUF
• Qwen/Qwen3.8-27B language:
• en
• th
• zh
• multilingual library_name: llama.cpp pipeline_tag: text-generation tags:
• gguf
• llama.cpp
• ptq1_0
• ternary
• abliterated
• uncensored
• lora-merge
• svd
• jev
• decision-model
• calibrated-probabilities
• qwen3_8
• prism
• bonsai

JEV Ternary Bonsai 2 27B — Abliterated + JEV Decision LoRA (PTQ1_0 GGUF)

This is BoldingBuilds' abliterated Ternary Bonsai 2 27B PTQ1_0 GGUF with the hotdogs JEV SVD LoRA merged in.

The JEV adapter is a decision-model LoRA: it pushes the model toward typed decisions, System-1/System-2 style reasoning paths, and calibrated probability output rather than free-form chat only. This release keeps the small PTQ1_0 ternary footprint while adding that decision behaviour.

Recommended usage: non-thinking mode. Run with --jinja --chat-template-kwargs '{"enable_thinking": false}' unless you intentionally increase the token budget for long internal reasoning.

Files in this repo

Requirements

This is a PrismML ternary GGUF. Use the PrismML llama.cpp fork rather than assuming mainline llama.cpp will load the PTQ1_0 pack.

git clone https://github.com/PrismML-Eng/llama.cpp
cmake -B build -DGGML_CUDA=ON
cmake --build build -j

CPU or Metal can also be used depending on your hardware; CUDA is the most common local-server path.

Usage

Direct answer mode, recommended

./build/bin/llama-server \
  -m jev-Ternary-Bonsai-2-27B-Abliterated-PTQ1_0.gguf \
  -ngl 99 \
  -c 8192 \
  --flash-attn on \
  --jinja \
  --chat-template-kwargs '{"enable_thinking": false}'

Then open http://127.0.0.1:8080.

Decision-style prompt

Decide: should we ship this release?
Evidence: tests pass, latency is within budget, one known edge case is documented.
Output a typed decision with a confidence interval and the main risk.

Non-thinking vs thinking mode

Use non-thinking mode for normal chat, direct answers, short decisions, and constrained max_tokens settings. It matches the style of the base abliteration evaluation and avoids spending the whole completion budget on internal reasoning before reaching the reply.

If you enable thinking, increase the token budget substantially and treat long reasoning as an experiment, not as the default serving configuration.

What is honest about this build

• The base model already removes much refusal behaviour; the JEV LoRA changes the style toward structured decisions and probabilities.
• The base-card refusal and HumanEval numbers describe the abliterated base before this JEV merge. This card does not claim fresh merged-model benchmark numbers.
• The LoRA is an exact adapter relative to stock Qwen3.8-27B, but merging into a ternary PTQ1_0 base is still a model-editing operation that should be evaluated by the user for their own task.
• This is not a safety guarantee. An abliterated model can comply with requests that a stock model would refuse.

Safety

Use this model under your own policy and applicable law. It is useful for research, red-teaming, local assistants, and decision experiments, but it should not be treated as moderated simply because it runs locally.

License and credits

• Base abliterated ternary release: BoldingBuilds/Ternary-Bonsai-2-27B-Abliterated-PTQ1_0-GGUF
• Ternary base and format: prism-ml/Ternary-Bonsai-2-27B-gguf
• Original model family: Qwen/Qwen3.8-27B
• Merged adapter: hotdogs/jev_qwen3.8-27B_svd_lora, derived from autotrust/JEV-27B-VL

License: Apache-2.0, matching the base release and the JEV LoRA card.

คู่มือภาษาไทยสั้นๆ

ใช้โมเดลนี้ผ่าน PrismML llama.cpp fork แล้วรันเซิร์ฟเวอร์ด้วย --jinja --chat-template-kwargs '{"enable_thinking": false}' สำหรับการใช้งานทั่วไป

เหตุผลคือ base abliteration ถูกทดสอบในโหมดปิด thinking และโหมดคิดยาวอาจใช้ token budget หมดก่อนตอบ ถ้าต้องการให้คิดยาวจริง ควรเพิ่ม token budget อย่างมีสติ ไม่ใช่เปิดทิ้งไว้เป็นค่าเริ่มต้น

This is a merged local research model: abliterated ternary Bonsai 2 plus JEV decision-model LoRA. It is not a leaderboard claim and not a safety guarantee.