This is a decensored version of TeichAI/Qwen3.6-27B-Claude-Opus-Reasoning-Distill-v2, made using Heretic v1.3.0
[!TIP] This model is reproducible!
See the README in the
reproducedirectory for more information.
Abliteration parameters
| Parameter | Value |
|---|---|
| direction_index | 41.77 |
| attn.o_proj.max_weight | 1.22 |
| attn.o_proj.max_weight_position | 51.32 |
| attn.o_proj.min_weight | 1.21 |
| attn.o_proj.min_weight_distance | 32.79 |
| mlp.down_proj.max_weight | 1.44 |
| mlp.down_proj.max_weight_position | 42.02 |
| mlp.down_proj.min_weight | 0.13 |
| mlp.down_proj.min_weight_distance | 37.15 |
Performance
| Metric | This model | Original model (TeichAI/Qwen3.6-27B-Claude-Opus-Reasoning-Distill-v2) |
|---|---|---|
| KL divergence | 0.0774 | 0 (by definition) |
| Refusals | 8/100 | 98/100 |
Qwen3.6 27B x Claude Opus 4.x - v2
Benchmarks
Qwen3.6-27B-Claude-Opus-Reasoning-Distill-v2
arc arc/e boolq hswag obkqa piqa wino
mxfp8 0.665,0.831,0.910,0.790,0.456,0.813,0.772
Qwen3.6-27B
arc arc/e boolq hswag obkqa piqa wino
mxfp8 0.647,0.803,0.910,0.773,0.450,0.806,0.742
Provided by @nightmedia. All benchmarks were done in mxfp8 precision
🧬 Datasets:
⚡ Use cases
- Coding
- Creative Writing
- Visual Understanding
- General Purpose
Citations and Contributions
- @unsloth - This qwen3 model was trained 2x faster with Unsloth and Huggingface's TRL library.
- @Qwen - Providing a fantastic, native-multimodal base model
Usage
If you need help setting up and configuring this model please follow the Qwen team's instructions in the original model's README