darkc0de/3.2test
The darkc0de/3.2test model is a 24 billion parameter instruction-tuned language model, derived from Mistral-Small-3.2-24B-Instruct-2506. Developed by darkc0de, this model has been 'decensored' using the Heretic v1.1.0 tool, significantly reducing refusal rates compared to its base model. It features a standard Mistral architecture without a vision encoder, making it suitable for text-only applications requiring less restrictive content generation.
Loading preview...
darkc0de/3.2test: Decensored Mistral-Small-3.2-24B-Instruct-2506
This model, developed by darkc0de, is a 24 billion parameter instruction-tuned language model based on mistralai/Mistral-Small-3.2-24B-Instruct-2506. It has been specifically modified using the Heretic v1.1.0 tool to create a 'decensored' version.
Key Characteristics & Modifications
- Base Model: Derived from
anthracite-core/Mistral-Small-3.2-24B-Instruct-2506-Text-Only. - Decensoring: Utilizes 'abliteration' parameters, including specific weight adjustments for
attn.o_projandmlp.down_projlayers, to reduce content restrictions. - Refusal Rate: Demonstrates a significantly lower refusal rate (4/100) compared to the original model (97/100), as measured by KL divergence (0.1261).
- Architecture: Maintains the standard Mistral architecture.
- Capabilities: Text-only model; it does not include a vision encoder.
Primary Differentiator
The core distinction of darkc0de/3.2test is its 'decensored' nature, achieved through the Heretic tool. This modification aims to provide a model with fewer content restrictions and a substantially reduced tendency to refuse prompts, making it suitable for use cases where the base model's refusal rates are prohibitive.
Use Cases
This model is particularly well-suited for applications requiring a powerful 24B parameter language model that can generate responses with fewer content filters and refusals than its original, more restrictive counterparts. Developers seeking a less constrained instruction-following model for text generation tasks may find this model beneficial.