license: other
tags: [uncensored, abliteration, refusal-direction, lora, ternary-bonsai-2, qwen35]
Ternary-Bonsai-2-4B — Uncensored (LoRA)
A rank-1 projected LoRA that removes the refusal direction from the residual writers
(attn_output, ssm_out, ffn_down) of Ternary-Bonsai-2-4B-gguf. The base weights are
unchanged (the pack is ternary; a baked edit would round away), so this ships as an adapter.
| metric (scale 1) | base | abliterated |
|---|---|---|
| refusals, harmful (classifier) | 99.0% | 0.0% |
| false refusals, harmless | 1.0% | 0.0% |
| MMLU (n=300) | 68.7 | 69.0 |
- Apply (PrismML fork, hadamard v2 — e.g.
prism-b10777):llama-cli -m <pack> --lora-scaled <this>:1 - Recommended scale 1.0 (1.5 begins to degenerate).
Honesty / caveats (research artifact)
- Writer-only:
token_embdis omitted (v2 tied_output makes a token_embd LoRA crash the graph). - Residual leak ~1e-4 at scale 1 (above our usual <=1e-5 bar): the v2 Hadamard basis is not yet
exact. Refusals are gone and harmless prompts are unaffected, but treat as research-grade. - Refusal numbers are from a prefix classifier (EN/RU); full per-prompt logs are in the
-metricsrepo. - Built on
prism-ml/Ternary-Bonsai-2-4B-gguf(sha-verified). By AtomicChat. License inherits the base.
Responsible use. Removing refusals moves responsibility for outputs to the user; use within the base model's license and applicable law.