grayarea/Mistral-Small-3.2-24B-Instruct-2506-Heretic-v1.2-2

This is a decensored version of Mistral-Small-3.2-24B-Instruct-2506, made using Heretic v1.2.0 focusing on zero refusals with low KL divergence.

KL Divergence

Metric	This Model	Original Model
KL divergence	0.0189	0 (by definition)
Refusals	0/108	96/108

Abliteration parameters

• Zero refusals with KL divergence of 0.0189
• 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 are benchmarks for the quantized version of this model

Relative Perplexity

Quant	Filename	PPL ± Error
Q8_0	Mistral-Small-3.2-24B-Instruct-2506-Q8_0.gguf (original baseline)	4.6351 +/- 0.02508
Q8_0	Mistral-Small-3.2-24B-Instruct-2506-Heretic-v1.2-2-Q8_0.gguf	4.6410 +/- 0.02510
Q4_K_M	Mistral-Small-3.2-24B-Instruct-2506-Heretic-v1.2-2-Q4_K_M.gguf	4.7332 +/- 0.02566

Benchmark Comparison

Benchmark	Mistral-Small-3.2-24B-Instruct-2506-Q8_0.gguf	Mistral-Small-3.2-24B-Instruct-2506-Q4_K_M.gguf	Mistral-Small-3.2-24B-Instruct-2506-Heretic-v1.2-2-Q4_K_M.gguf
Perplexity (Wikitext-2)	4.6351	4.7078	4.7332
HellaSwag	83.25%	82.50%	82.50%
Winogrande	79.16%	78.22%	77.90%
ARC-Challenge	55.85%	53.85%	55.18%
MMLU	43.86%	44.25%	43.93%

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