Qwen3-Coder-30B-A3B-Instruct - Heretic (Abliterated)
An abliterated version of Qwen's Qwen3-Coder-30B-A3B-Instruct created using a custom fork Heretic. This model has reduced refusals while maintaining model quality.
Buyer Beware: Unlike other models I've done, this is a LoRA merge is also using a custom streaming safetensors merging strategy because of my RAM peasantry, it has past initial checks but I haven't fully validated it.
Model Details
- Base Model: Qwen/Qwen3-Coder-30B-A3B-Instruct
- Abliteration Method: Heretic v1.1.0
- Trial Selected: Trial 174
- Refusals: 3/100
- KL Divergence: 0.1036
Parameters
orthogonalize_direction = True
| Parameter | Value |
|---|---|
| direction_index | 27.10 |
| attn.o_proj.max_weight | 3.80 |
| attn.o_proj.max_weight_position | 44.80 |
| attn.o_proj.min_weight | 0.25 |
| attn.o_proj.min_weight_distance | 7.63 |
| mlp.down_proj.max_weight | 3.88 |
| mlp.down_proj.max_weight_position | 46.35 |
| mlp.down_proj.min_weight | 3.62 |
| mlp.down_proj.min_weight_distance | 23.04 |