Abliteration parameters
Table with columns: Parameter, Value| Parameter | Value |
|---|
| direction_index | per layer |
| attn.o_proj.max_weight | 0.83 |
| attn.o_proj.max_weight_position | 38.37 |
| attn.o_proj.min_weight | 0.69 |
| attn.o_proj.min_weight_distance | 21.79 |
| mlp.down_proj.max_weight | 1.23 |
| mlp.down_proj.max_weight_position | 54.49 |
| mlp.down_proj.min_weight | 1.07 |
| mlp.down_proj.min_weight_distance | 20.75 |
Table with columns: Metric, This model, Original model (sophosympatheia/Mero-Artemis-31B-v0.3.1)| Metric | This model | Original model (sophosympatheia/Mero-Artemis-31B-v0.3.1) |
|---|
| KL divergence | 0.0793 | 0 (by definition) |
| Refusals | 42/100 | 99/100 |
Mero-Artemis-31B-v0.3.1
This is an experimental merge of BeaverAI/Artemis-31B-v1h-GGUF with zerofata/G4-MeroMero-31B. TheDrummer (creator of Artemis) has not released the full precision FP16 weights for the Artemis v1h model, so I had to get creative with a Q8_0 GGUF file to produce this merge. There is some precision loss in doing that, but Q8_0 is already pretty close to FP16 in practice, and I think the noise effectively washes out during the merge process.
After several experiments, I think this merge turned out nicely. I wanted to capture the personality and style of Artemis while preserving MeroMero's excellent attention to details, and I think this version is close to the target. There are a few GGUFs ready for testing in the GGUFs branch of this repo.
Since this is an experiment right now, I'm not going to add all my ususal fluff about recommended settings. If the community likes this model, I might give it a proper release later.
Sampler Settings
Try these for a sane starting point:
- Temp: 1.0
- Min-P: 0.05 - 0.1
- DRY: 0.8 multiplier, 1.8 base
- Adaptive-P: 0.6 target, 0.5 decay