mlabonne/NeuralLlama-3-8B-Instruct-abliterated

Model Overview

mlabonne/NeuralLlama-3-8B-Instruct-abliterated is an experimental 8 billion parameter Llama 3 Instruct model, fine-tuned with Direct Preference Optimization (DPO) on the mlabonne/orpo-dpo-mix-40k dataset. This fine-tuning process aims to enhance the model's performance, specifically recovering MMLU scores that were reduced during the initial 'abliteration' of the base Llama 3 8B Instruct model.

Key Characteristics

• Uncensored Nature: The model is designed to be uncensored, offering flexibility for various applications where strict content alignment is not a primary concern.
• Performance Improvement: Evaluation on the Open LLM Leaderboard and Nous benchmarks indicates that this DPO fine-tune improves upon the abliterated source model, surpassing the original Meta-Llama-3-8B-Instruct in average scores on the Nous benchmark.
• Llama 3 Architecture: Built upon the Llama 3 8B Instruct foundation, it retains the core capabilities of the Llama family.

Use Cases

• Role-playing: Its uncensored nature makes it particularly well-suited for creative and open-ended role-playing scenarios.
• Applications without Alignment Requirements: Ideal for use cases where the absence of strict content filters is beneficial or necessary.

Evaluation Highlights

On the Nous benchmark, mlabonne/Llama-3-8B-Instruct-abliterated-dpomix achieved an average score of 52.26, outperforming meta-llama/Meta-Llama-3-8B-Instruct (51.34) and the failspy/Meta-Llama-3-8B-Instruct-abliterated-v3 (51.21) across metrics like AGIEval, GPT4All, TruthfulQA, and Bigbench.