Overview

huihui-ai/Qwen2.5-7B-Instruct-abliterated-v2 is an uncensored variant of the Qwen/Qwen2.5-7B-Instruct model, developed by huihui-ai. This 7.6 billion parameter model leverages an 'abliteration' technique, as detailed in a Hugging Face article, to modify its response filtering. It builds upon a previous iteration, offering improved performance.

Key Capabilities & Performance

This model is designed to provide instruction-following capabilities without the inherent censorship present in its base model. Evaluations show its performance across several benchmarks:

• IF_Eval: Achieves 77.82, outperforming both the original Qwen2.5-7B-Instruct and the prior abliterated version.
• MMLU Pro: Scores 42.03.
• TruthfulQA: Scores 57.81.
• BBH: Scores 53.01.
• GPQA: Achieves 32.17, slightly surpassing the original model.

Usage Considerations

This model is particularly suited for use cases where unfiltered or uncensored responses are required, due to its abliterated nature. Developers can integrate it using the Hugging Face transformers library, with provided Python code examples for loading and interaction. The model supports a large context window of 131,072 tokens, enabling extensive conversational or document-based applications.