huihui-ai/Qwen2.5-7B-Instruct-abliterated-v3
The huihui-ai/Qwen2.5-7B-Instruct-abliterated-v3 is a 7.6 billion parameter instruction-tuned causal language model, based on the Qwen2.5-7B-Instruct architecture. Developed by huihui-ai, this model has been modified using 'abliteration' techniques to reduce refusal behaviors, offering an uncensored response profile. It supports a 131072-token context length and is designed for general text generation tasks where reduced content moderation is desired. The model is multilingual, supporting languages such as Chinese, English, French, Spanish, and more.
Loading preview...
Model Overview
The huihui-ai/Qwen2.5-7B-Instruct-abliterated-v3 is an instruction-tuned large language model with 7.6 billion parameters, derived from the Qwen/Qwen2.5-7B-Instruct base model. Its primary distinction lies in its "abliterated" nature, meaning it has undergone a process to significantly reduce refusal behaviors and content moderation, resulting in an uncensored output profile. This modification was implemented using techniques described in the remove-refusals-with-transformers project, serving as a proof-of-concept for removing refusals without relying on TransformerLens.
Key Characteristics
- Uncensored Output: Engineered to minimize refusals and provide direct responses, even to potentially sensitive queries.
- Multilingual Support: Capable of processing and generating text in numerous languages, including Chinese, English, French, Spanish, Portuguese, German, Italian, Russian, Japanese, Korean, Vietnamese, Thai, and Arabic.
- Extended Context Window: Features a substantial context length of 131072 tokens, allowing for processing longer inputs and maintaining conversational coherence over extended interactions.
Performance and Use Cases
While the abliteration process aims to reduce refusals, evaluation benchmarks indicate some trade-offs in general performance compared to the original Qwen2.5-7B-Instruct. For instance, it shows slightly lower scores on MMLU Pro, TruthfulQA, and BBH. However, it offers an alternative for applications requiring less restrictive content filtering. This model is suitable for developers and researchers exploring methods of controlling model behavior and for use cases where an uncensored, direct response style is preferred, provided the slight performance variations are acceptable.
Top 3 parameter combinations used by Featherless users for this model. Click a tab to see each config.