Overview

saidutta69/Qwen2.5-3B-Instruct-heretic is a 3.09 billion parameter instruction-tuned causal language model, built upon the Qwen2.5 architecture. This particular version is a decensored variant of the original Qwen/Qwen2.5-3B-Instruct, created using the Heretic v1.2.0 tool. It aims to provide a more permissive and less restrictive conversational experience.

Key Capabilities

• Decensored Responses: Significantly reduces refusals, with only 2 refusals out of 100 prompts compared to 96/100 for the base model.
• Enhanced Knowledge: Inherits Qwen2.5's improvements in coding and mathematics due to specialized expert models.
• Improved Instruction Following: Better at understanding and executing complex instructions.
• Long-Text Generation: Capable of generating long texts, up to 8K tokens, within a 32K token context window.
• Structured Data & Output: Improved understanding of structured data (e.g., tables) and generation of structured outputs, particularly JSON.
• Multilingual Support: Supports over 29 languages, including major global languages.

Good for

• Applications requiring an uncensored or less restrictive conversational AI.
• Use cases where the original Qwen2.5-3B-Instruct's refusal rate is too high.
• Tasks involving coding, mathematics, and complex instruction following where a smaller, efficient model is preferred.
• Generating long, coherent texts or structured outputs like JSON.