Overview
This model, soaring0616/Qwen2.5-7B-Instruct-heretic, is a 7.61 billion parameter instruction-tuned causal language model. It is a decensored version of the original Qwen/Qwen2.5-7B-Instruct model, created using the Heretic v1.0.1 tool. The decensoring process significantly alters the model's refusal behavior, reducing refusals from 95/100 in the original to 43/100 in this variant, as measured by KL divergence of 0.19.
Key Capabilities
- Reduced Refusals: Demonstrates a notable decrease in refusal rates compared to the base Qwen2.5-7B-Instruct model.
- Enhanced Instruction Following: Improved ability to adhere to diverse instructions and system prompts.
- Long Context Support: Features a substantial context length of 131,072 tokens, with generation capabilities up to 8,192 tokens.
- Multilingual: Supports over 29 languages, including major global languages like Chinese, English, French, Spanish, German, and Japanese.
- Structured Output: Better at understanding structured data (e.g., tables) and generating structured outputs like JSON.
- Improved Core Abilities: Inherits Qwen2.5's advancements in coding, mathematics, and knowledge.
Good For
- Applications requiring a less restrictive instruction-following model.
- Tasks involving long-form text generation and comprehension.
- Use cases demanding structured data processing and output generation.
- Multilingual applications across a broad range of languages.