HERETICODER-2.5-7B-IT: Decensored Code Generation

Cannae-AI/HERETICODER-2.5-7B-IT is a 7.6 billion parameter instruction-tuned model derived from the Qwen2.5-Coder-7B-Instruct base. Its primary distinction lies in its "abliterated" and "decensored" nature, achieved through the Heretic process, which significantly reduces model refusals.

Key Capabilities

• Decensored Output: Engineered to minimize refusals, with a reported 1/100 refusal rate, making it highly permissive for code-related queries.
• Zero KL Divergence: Maintains a 0.00 KL divergence, indicating high fidelity to its base model's underlying knowledge while removing censorship.
• Extended Context Window: Features a substantial 131,072 token context length, enabling it to process and generate extensive code snippets and complex technical instructions.
• Code-Optimized: Inherits the code generation capabilities of its Qwen2.5-Coder base, making it proficient in programming tasks.

Good For

• Applications requiring uncensored or less restrictive code generation.
• Developers needing a model that avoids common refusal patterns in technical contexts.
• Handling large codebases or complex programming problems due to its extended context window.