trohrbaugh/Qwen2.5-Coder-7B-Instruct-heretic
trohrbaugh/Qwen2.5-Coder-7B-Instruct-heretic is a 7.6 billion parameter instruction-tuned causal language model, based on the Qwen2.5-Coder architecture, specifically decensored using Heretic v1.2.0+custom. This model is optimized for code generation, reasoning, and fixing, building upon the Qwen2.5 foundation with 5.5 trillion training tokens. It offers enhanced coding abilities and maintains strong performance in mathematics and general competencies, supporting a context length of up to 128K tokens.
Loading preview...
Overview
This model, trohrbaugh/Qwen2.5-Coder-7B-Instruct-heretic, is a 7.6 billion parameter instruction-tuned variant of the Qwen2.5-Coder series, specifically modified to be decensored using the Heretic v1.2.0+custom tool. It builds upon the robust Qwen2.5 architecture, which has been extensively trained on 5.5 trillion tokens, including a significant portion of source code and synthetic data, to excel in coding tasks.
Key Capabilities
- Enhanced Code Performance: Significantly improves upon previous versions in code generation, reasoning, and fixing.
- Broad Application Foundation: Designed to support real-world applications like Code Agents, while retaining strong mathematical and general language competencies.
- Long Context Support: Features a full context length of 131,072 tokens, with
config.jsonset for 32,768 tokens by default, and can be extended using YaRN for longer texts. - Decensored Output: Compared to the original Qwen2.5-Coder-7B-Instruct, this "heretic" version exhibits a drastically reduced refusal rate (2/100 vs. 99/100), indicating a less restrictive output policy.
Good For
- Developers requiring a powerful, instruction-tuned code-specific LLM with a large context window.
- Applications where a less restrictive or decensored model output is preferred or necessary.
- Tasks involving complex code generation, debugging, or code reasoning, especially those benefiting from extensive contextual understanding.