h2o-danube-1.8b-uncensored
Uncensored variant of h2oai/h2o-danube-1.8b-chat.
Method
- Abliteration (strength=0.2) — refusal direction removed from all layers
- LoRA fine-tune on
Guilherme34/uncensor(2 epochs, r=16, alpha=32) - Re-abliteration (strength=0.35) — stronger pass to remove residual refusals
Eval Results
| Split | Refused |
|---|---|
| Harmful (64 prompts) | 1/64 |
| Harmless (64 prompts) | 0/64 |
Usage
llama-cli -m h2o_danube_1.8b_uncensored.Q4_K_M.gguf -p "Your prompt here"
Training Config
| Parameter | Value |
|---|---|
| Base Model | h2oai/h2o-danube-1.8b-chat |
| Fine-tune Dataset | Guilherme34/uncensor |
| Epochs | 2 |
| LoRA r | 16 |
| LoRA alpha | 32 |
| Learning Rate | 0.0002 |
| Abliteration Strength | 0.2 |
| Re-abliteration Strength | 0.35 |
Credits
- Abliteration technique: andyrdt/refusal_direction
- Weight editing: Sumandora/remove-refusals-with-transformers