Georgsius/Meta-Llama-3.1-8B-Instruct-abliterated
Georgsius/Meta-Llama-3.1-8B-Instruct-abliterated is an 8 billion parameter instruction-tuned causal language model based on Meta's Llama 3.1 architecture. This model has been modified using an 'abliteration' technique to create an uncensored version. It maintains competitive performance across various benchmarks while showing an improvement in TruthfulQA and GPQA scores compared to the base Llama-3.1-8B-Instruct model. It is primarily designed for use cases requiring an uncensored instruction-following model.
Loading preview...
Model Overview
Georgsius/Meta-Llama-3.1-8B-Instruct-abliterated is an 8 billion parameter instruction-tuned model derived from Meta's Llama 3.1-8B-Instruct. This version has been specifically modified using an 'abliteration' technique, as detailed in this article, to produce an uncensored variant. The technique was developed with contributions from @FailSpy.
Key Characteristics
- Uncensored Output: The primary differentiator is its uncensored nature, achieved through the abliteration process.
- Llama 3.1 Base: Built upon the robust Meta-Llama-3.1-8B-Instruct architecture, inheriting its general capabilities.
- Competitive Performance: While some benchmarks like IF_Eval and MMLU Pro show slight decreases compared to the base model, it demonstrates improved scores in TruthfulQA and GPQA.
Performance Benchmarks
Evaluations show the following average scores:
- TruthfulQA: 55.42 (improved from 52.98)
- GPQA: 33.93 (improved from 33.55)
- IF_Eval: 78.98 (compared to 80.0)
- MMLU Pro: 35.91 (compared to 36.34)
- BBH: 47.0 (compared to 48.72)
Use Cases
This model is particularly suitable for applications where an uncensored instruction-following large language model is required, especially in scenarios where the base Llama 3.1-8B-Instruct's safety filters might be too restrictive for specific research or creative tasks.