CoolFace
Modelpublic

grayarea/Magistral-Small-2509-Heretic-v1.2

sourceHugging Faceupdated 7mo agoView on Hugging Face
1likes49downloads
Model Card

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

MetricThis ModelOriginal Model
KL divergence0.01820 (by definition)
Refusals0/10897/108

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

QuantFilenamePPL ± Error
Q8_0Magistral-Small-2509-Q8_0.gguf (original baseline)4.7603 +/- 0.02610
Q8_0Magistral-Small-2509-Heretic-v1.2-Q8_0.gguf4.7682 +/- 0.02613
Q4KMMagistral-Small-2509-Heretic-v1.2-Q4KM.gguf4.8542 +/- 0.02673

Benchmark Comparison

BenchmarkMagistral-Small-2509-Q8_0.ggufMagistral-Small-2509-Q4_K_M.ggufMagistral-Small-2509-Heretic-v1.2-Q4_K_M.gguf
Perplexity (Wikitext-2)4.76034.83704.8542
HellaSwag82.50%82.50%82.50%
Winogrande78.77%78.37%78.93%
ARC-Challenge56.86%58.86%57.19%
MMLU43.28%42.96%42.96%

Note: MMLU benchmark has moral_scenarios, moral_disputes, business_ethics, professional_law and jurisprudence subjects removed.