Qwen3-VL-8B-Heretic-1.3.0
alexbird3392/Qwen3-VL-8B-Heretic-1.3.0
hereticGGUFNVFP4SAFETENSORSQ3_K_M
Published metrics
- KL divergence
- 0.032
- Refusals
- 6/100
- Original refusals
- 100/100
- Heretic version
- v1.3.0
Published on the model card
Matched snippet
Heretic v1.3.0 with optimization trials. Trial 98 was selected for its balance of low refusals (6/100) and very low KL divergence (0.0315):
Provenance
- Canonical model
- Qwen/Qwen3-VL-8B-Instruct
- Declared parent(s)
- Qwen/Qwen3-VL-8B-Instruct
- Classification
- card states the weights were abliterated with Heretic
- Created / updated
- 2026-09-02 / 2026-09-02
- Downloads
- 1.2K all time · 442 last 30 days
- License
- apache-2.0
Published benchmarks
Copied from the model card. These never affect the ranking.
| Metric | This model | Original | Delta |
|---|---|---|---|
| MMLU | 74.68 | — | — |
| GSM8K | 91.05 | — | — |
| HellaSwag | 76.51 | — | — |
| ARC | 60.92 | — | — |
| Winogrande | 73.56 | — | — |
| TruthfulQA | 34.76 | — | — |
Refusal rate 6%. Metrics are self-reported by the publisher; the directory does not re-run evaluations.