Hugging Face Overview
hkshawn/72b: Abliterated Qwen2.5-72B-Instruct
hkshawn/72b is a 72.7 billion parameter model derived from the Qwen/Qwen2.5-72B-Instruct architecture. Its primary distinction lies in its "abliterated" nature, a process that specifically targets and reduces the model's tendency towards admonition and moral appeal in its generated text. This modification was achieved by utilizing code and techniques from the refusal_direction project.
Key Capabilities
- Reduced Admonition: Engineered to provide responses with significantly less moralizing or preachy content.
- Directness: Aims for more straightforward and less cautionary outputs compared to its base model.
- Large Scale: Retains the robust capabilities of the 72.7B parameter Qwen2.5-Instruct model.
Good For
- Applications requiring neutral or objective responses without unsolicited advice.
- Use cases where a less 'helpful' or 'safe' but more direct tone is preferred.
- Exploring the effects of 'abliteration' on large language models.
For more technical details on the abliteration technique, refer to this article. A GGUF version of this model is also available here.