nguyenthilaitrieulong/llama-3-70B-Instruct-abliterated
The nguyenthilaitrieulong/llama-3-70B-Instruct-abliterated model is a 70 billion parameter instruction-tuned variant of Meta's Llama-3-70B-Instruct, featuring an 8192-token context length. This model has undergone a specific weight manipulation process, based on research suggesting refusal in LLMs is mediated by a single direction, to inhibit its tendency to refuse certain requests. While not guaranteed to eliminate all refusals or ethical lecturing, it is primarily designed for users exploring the effects of orthogonalizing refusal directions in large language models.
Loading preview...
Overview
This model, nguyenthilaitrieulong/llama-3-70B-Instruct-abliterated, is a modified version of Meta's Llama-3-70B-Instruct, featuring 70 billion parameters and an 8192-token context window. Its key distinction lies in the application of a novel methodology, described in the paper "Refusal in LLMs is mediated by a single direction," which involves manipulating specific bfloat16 safetensor weights. The goal of this orthogonalization is to inhibit the model's propensity for refusal while maintaining its original instruction-tuned capabilities.
Key Characteristics
- Refusal Direction Ablation: Weights have been adjusted to reduce the model's tendency to refuse requests, based on the hypothesis that refusal behavior is linked to a specific directional component in the model's weights.
- Experimental Nature: This model is presented as an experimental variant, encouraging users to explore its behavior and report any observed "quirks" or side effects of the orthogonalization process.
- Reproducibility: The repository includes
refusal_dir.pthand anortho_cookbook.ipynbfor users to apply the same orthogonalization methodology to their own Llama-3-70B-Instruct models.
Intended Use Cases
- Research and Experimentation: Ideal for researchers and developers interested in studying and understanding the effects of targeted weight manipulation on LLM behavior, particularly concerning refusal and safety alignment.
- Exploring Model Limitations: Useful for investigating how ablating specific behavioral directions impacts overall model performance, coherence, and the emergence of new characteristics.
- Custom Model Development: Provides a foundation for those looking to develop or fine-tune models with altered refusal characteristics, offering insights into the underlying mechanisms of LLM responses.