Model Overview

grayarea/Magistral-Small-2509-Heretic-v1.2 is a 24 billion parameter language model based on Magistral-Small-2509, distinguished by its "decensored" nature. This version was created using Heretic v1.2.0, focusing on achieving zero refusals with a minimal impact on the model's original characteristics, as indicated by a low KL divergence.

Key Characteristics

• Zero Refusals: The model demonstrates 0 refusals out of 108 test cases, significantly down from 97/108 in the original model.
• Low KL Divergence: Achieves a KL divergence of 0.0182 compared to the original model, suggesting a high degree of fidelity to the base model's knowledge and capabilities.
• Abliteration Parameters: Utilizes a custom Heretic training dataset and configuration, employing Magnitude-Preserving Orthogonal Ablation (MPOA), full row renormalization, and Winsorization Quantile 0.997.
• Performance: Maintains relative perplexity close to the original model, with a Q8_0 quantized version showing 4.7682 +/- 0.02613 compared to the original's 4.7603 +/- 0.02610. General benchmarks like HellaSwag, Winogrande, and ARC-Challenge also show comparable results.

Use Cases

This model is particularly well-suited for applications where:

• Unrestricted Content Generation: The primary requirement is to generate responses without content-based refusals.
• Maintaining Original Model Performance: Users need the core capabilities of Magistral-Small-2509 but with the added flexibility of a decensored output.
• Research into Model Alignment: It can be a valuable tool for studying the effects of alignment techniques and their impact on model behavior and refusal rates.