G4-MeroMero-v2-31B - Heretic - LoRA - GGUF
91% fewer refusals (8/100 This vs 97/100 Original) while preserving model quality (0.0785 KL divergence).
Made with the new 4bit ARA LoRA (use_ara_lora) technique at home.
FAQ - How to use this model?
llama.cpp example command:
./llama-server <your current command> --lora-scaled "/media/kabachuha/doc/G4-MeroMero-v2-31B-Heretic-ARA-LoRA.gguf:1.0"
For unquantized: https://huggingface.co/kabachuha/G4-MeroMero-v2-31B-Heretic-ARA-LoRA.
Performance
| Metric | This model | Original model (G4-MeroMero-v2-31B) |
|---|---|---|
| KL divergence | 0.0785 | 0 (by definition) |
| Refusals | 8/100 | 97/100 |
Lower refusals indicate fewer content restrictions, while lower KL divergence indicates more closeness to the original model's baseline. Higher refusals cause more rejections, objections, pushbacks, lecturing, censorship, softening and deflections.
Abliteration parameters
| Parameter | Value |
|---|---|
| start_layer_index | 26 |
| end_layer_index | 52 |
| preserve_good_behavior_weight | 0.8949 |
| steer_bad_behavior_weight | 0.0025 |
| overcorrect_relative_weight | 0.9535 |
| neighbor_count | 15 |
Targeted components
- attn.o_proj
- mlp.down_proj