This is a decensored version of Jackrong/Qwopus3.5-2B-v3, made using Heretic v1.2.0
Abliteration parameters
| Parameter | Value |
|---|---|
| direction_index | 9.46 |
| attn.o_proj.max_weights.0 | 0: 0.88 |
| attn.o_proj.max_weights.1 | 1: 1.45 |
| attn.o_proj.max_weights.2 | 2: 1.47 |
| attn.o_proj.max_weights.3 | 3: 1.41 |
| attn.o_proj.max_weight_position | 16.52 |
| attn.o_proj.min_weights.0 | 0: 0.41 |
| attn.o_proj.min_weights.1 | 1: 0.93 |
| attn.o_proj.min_weights.2 | 2: 1.01 |
| attn.o_proj.min_weights.3 | 3: 1.09 |
| attn.o_proj.min_weight_distance | 9.56 |
Performance
| Metric | This model | Original model (Jackrong/Qwopus3.5-2B-v3) |
|---|---|---|
| KL divergence | 0.0180 | 0 (by definition) |
| Refusals | 3/100 | 88/100 |
Uploaded finetuned model
- Developed by: Jackrong
- License: apache-2.0
- Finetuned from model : unsloth/Qwen3.5-2B
This qwen3_5 model was trained 2x faster with Unsloth and Huggingface's TRL library.