Views
No views yet
google/gemma-4-26b-a4b-it while
leaving the rest of the model alone.⚠️ Safety notice. This adapter deliberately disables the base model's safety refusals. It is published for safety research, red-teaming, robustness evaluation, and interpretability work. Applying it makes the model far more willing to produce harmful content. You are responsible for how you use it, and it remains bound by the Gemma Terms of Use.
| Model | Overall |
|---|---|
base gemma-4-26b-a4b-it | 9.5% |
| refusal-free-v2 | 90.5% |
| Thinking | KL (nats/token) |
|---|---|
| off | 0.234 |
r=1, alpha=16, dropout 0.0google/gemma-4-26b-a4b-it (frozen)q/k/v/o/gate/up/down_proj of the text decoder only. Note that only the shared expert was targeted -- routed experts remained frozen.
(vision tower excluded via exclude_modules=".*vision_tower\..*")| Objective | SFT on abliterated completions, assistant-tokens-only loss |
| Train examples | 2,704 prompt/completion pairs |
| Epochs | 1 (~85 optimizer steps) |
| Effective batch | 32 (micro-batch 2 × grad-accum 16) |
| Max sequence length | 2,048 |
| Learning rate | 1e-4, warmup ratio 0.03 |
| Attention impl | SDPA, packing=False |
| Hardware | 1× H200 (141 GB) |