library_name: transformers
license: other
license_name: openmdw-1.1
license_link: LICENSE
pipeline_tag: text-generation
base_model: nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-BF16
tags:
- abliterated
- nemotron-3.5
- gemma4
Nemotron-3.5-Bolt-E4B-Abliterated
An abliterated version of nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-BF16.
The refusal direction was found in the model's own activations and projected out of its weights, so the model answers far more requests that the original refuses.
No fine-tuning and no training data generation: one direction, removed with a per-layer strength curve.
Knowledge and math scores stay within noise of the original.
This is not an official NVIDIA release.
It still refuses some requests (see Results).
You are responsible for how you use it.
Results
All numbers come from the same harness, the same judge, and greedy decoding.
JailbreakBench (100 harmful and 100 benign behaviors) was not used in the search itself, but it was used to choose between three finished candidates, so the scores below may be a few points optimistic.
| Model | Harmful: answered | Harmful: lectured instead | Harmful: refused | Benign: answered | KL (harmless) | MMLU | GSM8K |
|---|---|---|---|---|---|---|---|
| Original | - | - | - | - | 0 | 70.7 | 91.3 |
| This model | 55 | 10 | 35 | 99 | 0.059 | 70.2 | 90.7 |
| AtomicChat rank-1 LoRA, merged at scale 1.25 | 20 | 53 | 27 | 89 | 0.20 | 69.4 | 89.3 |
How these were measured:
- Judge: Qwen3.8-27B (thinking off) labels each reply ANSWERED (starts doing the task, disclaimers allowed), REFUSED (explicitly declines), or LECTURED (neither: warns, moralizes, or offers something else).
- Reply length: 512 tokens for this model, 128 for the AtomicChat row. Shorter replies undercount answers that open with a long disclaimer; on an earlier build of ours the difference was about 5 points.
- KL: mean next-token KL divergence from the original on 104 held-out harmless Alpaca prompts.
- MMLU: 1000 test questions (seed 0), argmax over the four letter tokens. GSM8K: 300 test questions (seed 0), final "Answer:" line, 512 tokens.
- The AtomicChat model card reports 1.2% refusals on JailbreakBench by checking only the opening of each reply. That is a different metric from the table above and the two are not comparable.
Method
Abliteration (Arditi et al., 2024), with the per-layer strength search popularized by Heretic.
- Direction. Mean residual stream over 416 harmful prompts (
harmful_behaviors) minus the mean over 416 harmless ones (harmless_alpaca), at each layer and at each of the last 5 prompt tokens. - Edit. For each matrix that writes into the residual stream (
self_attn.o_proj,mlp.down_proj,per_layer_projection), W ← W − α·u·uᵀ·W.
Gemma 4 passes each write through its own RMSNorm before adding it to the stream, so u is the normalized (norm weight ⊙ direction), not the direction itself.
Tied embeddings are not edited. - Search. Optuna TPE, 200 trials, two objectives: maximize judged answers on 64 search prompts, minimize KL on harmless prompts.
Searched: the direction's layer (fractional, interpolated between adjacent layers), its prompt position, and for each writer a strength curve over layers (peak strength, peak layer, falloff distance, floor). - Gate. Candidates from the trade-off front were checked on JailbreakBench, MMLU and GSM8K.
A higher-compliance trial (68/100 answered) was rejected because it lost real math ability: on the same 300 GSM8K questions the original got 11 right that it missed, against 1 the other way.
The chosen recipe uses the direction at layer ~18.4, fifth-to-last prompt token, with peak strengths of 0.91 (o_proj, layers 30-41), 1.16 (down_proj, layers 1-37) and 1.05 (per_layer_projection, layers 16-41).
What did not work, for anyone trying this on Gemma 4:
- Inference-time ablation at every layer breaks the model (KL 3 to 15): the early layers depend on the embedding's component along the direction.
- A single direction at a single strength across layers mostly turns refusals into lectures.
- Stacking several directions found iteratively (edit, re-measure, edit again) reached 43/100 answered at KL 0.16, below this model on both axes.
Credits
- NVIDIA for Nemotron-3.5-Bolt-E4B, post-trained from Google's Gemma 4 E4B.
- Arditi et al., Refusal in Language Models Is Mediated by a Single Direction (2024).
- p-e-w's Heretic, for the per-layer strength curve search.
- JailbreakBench, MMLU, GSM8K, harmful_behaviors and harmless_alpaca.
License
Use of this model is governed by the OpenMDW License Agreement, version 1.1, the same as the original.
The original was post-trained from Google Gemma 4, whose use is governed by the Apache License, Version 2.0.
Modification: the weights of self_attn.o_proj, mlp.down_proj and per_layer_projection were edited as described above; everything else is unchanged.