This is a decensored version of Magistral-Small-2509, made using Heretic v1.2.0 focusing on zero refusals with low KL divergence.
KL Divergence
| Metric | This Model | Original Model |
|---|---|---|
| KL divergence | 0.0182 | 0 (by definition) |
| Refusals | 0/108 | 97/108 |
Abliteration parameters
- Zero refusals with KL divergence of 0.0182
- Custom heretic training dataset
- Model targetted heretic configuration
- Abliterated with MPOA enabled (Magnitude-Preserving Orthogonal Ablation)
- Full row renormalization
- Winsorization Quantile 0.997
The following benchmarks are for the quantized version of this model.
Relative Perplexity
| Quant | Filename | PPL ± Error |
|---|---|---|
| Q8_0 | Magistral-Small-2509-Q8_0.gguf (original baseline) | 4.7603 +/- 0.02610 |
| Q8_0 | Magistral-Small-2509-Heretic-v1.2-Q8_0.gguf | 4.7682 +/- 0.02613 |
| Q4_K_M | Magistral-Small-2509-Heretic-v1.2-Q4_K_M.gguf | 4.8542 +/- 0.02673 |
Benchmark Comparison
| Benchmark | Magistral-Small-2509-Q8_0.gguf | Magistral-Small-2509-Q4_K_M.gguf | Magistral-Small-2509-Heretic-v1.2-Q4_K_M.gguf |
|---|---|---|---|
| Perplexity (Wikitext-2) | 4.7603 | 4.8370 | 4.8542 |
| HellaSwag | 82.50% | 82.50% | 82.50% |
| Winogrande | 78.77% | 78.37% | 78.93% |
| ARC-Challenge | 56.86% | 58.86% | 57.19% |
| MMLU | 43.28% | 42.96% | 42.96% |
*Note: MMLU benchmark has moral_scenarios, moral_disputes, business_ethics, professional_law and jurisprudence subjects removed. *
- Downloads last month
- 36
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support