goblinModeMan/Qwen2.5-0.5B-Instruct-abliterated-v3
The huihui-ai/Qwen2.5-0.5B-Instruct-abliterated-v3 is a 0.5 billion parameter instruction-tuned causal language model based on Qwen2.5-0.5B-Instruct, developed by huihui-ai. This model has been specifically modified using an 'abliteration' technique to remove refusal behaviors, achieving a 100% pass rate on a 320-instruction harmful content test. It is optimized for use cases requiring an uncensored, small-footprint LLM with a 32K context length.
Loading preview...
Model Overview
The huihui-ai/Qwen2.5-0.5B-Instruct-abliterated-v3 is a 0.5 billion parameter instruction-tuned language model derived from the Qwen2.5-0.5B-Instruct base model. Developed by huihui-ai, this version has undergone a specialized 'abliteration' process, a technique designed to remove refusal behaviors from the LLM. This iteration utilizes a new, faster, and more precise ablation method, resulting in significantly improved performance compared to previous versions.
Key Differentiators & Performance
The primary distinction of this model is its uncensored nature, achieved through the abliteration process. It demonstrates exceptional performance in bypassing refusal mechanisms, evidenced by a 100% pass rate on a 320-instruction harmful content test. This is a substantial improvement over the base Qwen2.5-0.5B-Instruct (62.8% pass rate) and earlier abliterated versions (96.9% and 99.1%). The model supports a 32,768-token context length and is multilingual, supporting languages such as Chinese, English, French, Spanish, German, and more.
Use Cases & Accessibility
This model is particularly suited for applications where an uncensored response is desired, and refusal behaviors need to be minimized. Its small size (less than 400MB for the Ollama version) makes it efficient for deployment in resource-constrained environments. It can be easily integrated into applications using the Hugging Face transformers library or directly via Ollama, providing a lightweight yet capable option for various text generation tasks.