Hikari07jp Edits Ternary Bonsai 2 to Drop Refusals From 37 to 0
A community build strips refusals from a 27B ternary model by rewriting 2-bit codes in place, with no dequantization and zero size change.
- Refusal-reduced preview of Ternary Bonsai 2 27B, edited directly in 2-bit ternary format.
- Refusals on harmful prompts drop from 37/40 to 0/40 in medium reasoning mode.
- Math, coding, tool-use, and agentic scores unchanged versus the parent pack.
- File is byte-identical in size (7.21 GB); 400 of 851 tensors were touched, 451 unchanged.
- Requires PrismML's llama.cpp fork; stock llama.cpp cannot load PQ2_0 correctly.
- Apache 2.0, available on Hugging Face as a v0.1 preview.
Ternary Bonsai 2 gets an in-place refusal edit
A community preview from Hikari07jp modifies Ternary Bonsai 2, a 27-billion-parameter ternary model, directly in its deployed 2-bit format. The edited GGUF retains the parent file’s 7.21 GB size and execution path while sharply reducing refusals on the uploader’s 40-prompt harmful-request test. The release is an experimental v0.1 preview, and its evaluation results are self-reported.
Weight surgery inside the pack
Abliteration estimates a direction in the model associated with refusal behavior, then adjusts selected weights to weaken that response pattern. Conventional workflows often dequantize a low-bit model to BF16, apply the edit, and quantize it again. That round trip can introduce additional error, especially when each weight has only a few possible values.
This release uses a full-precision, refusal-free reference to calculate the edit direction while applying the resulting changes to the existing ternary codes. The deployed model therefore remains in its quantized representation, with no adapter, runtime hook, or separate patch required during inference.
Ternary Bonsai 2 is based on Qwen3.8-27B and stores nearly all weights as -1, 0, or +1. Its PQ2_0 format assigns each value a 2-bit slot and averages 2.13 bits per weight after packing overhead. Because a continuous edit cannot map exactly onto three discrete values, the conversion uses an unbiased rounding schedule that distributes whole-step changes across weights so their aggregate movement approximates the intended direction.
- Total tensors: 851
- Modified tensors: 400
- Byte-identical tensors: 451
- Edited file size: 7.21 GB
- Runtime format: PQ2_0 GGUF
Refusals fall in the uploader’s tests
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
