mlabonne/Meta-Llama-3.1-8B-Instruct-abliterated

Overview

mlabonne/Meta-Llama-3.1-8B-Instruct-abliterated is an 8 billion parameter instruction-tuned model derived from Meta's Llama 3.1 architecture. Its primary distinction is being an uncensored version, achieved through a technique called abliteration. This process modifies the model to remove certain content restrictions, offering greater flexibility in generated responses.

Key Characteristics

• Uncensored Output: Designed to provide less restricted content generation compared to its base model.
• Abliteration Technique: Created using a specific method to modify model behavior, as detailed in this article.
• Llama 3.1 Base: Leverages the foundational capabilities of the Meta Llama 3.1 8B Instruct model.
• Context Length: Supports a substantial context window of 32,768 tokens.

Performance Insights

Evaluations on the Open LLM Leaderboard show an average score of 23.13. Specific metrics include:

• IFEval (0-Shot): 73.29
• BBH (3-Shot): 27.13
• MMLU-PRO (5-shot): 27.81

Quantizations Available

Various quantized versions are provided for optimized deployment:

• New GGUF: mlabonne/Meta-Llama-3.1-8B-Instruct-abliterated-GGUF
• ZeroWw GGUF: ZeroWw/Meta-Llama-3.1-8B-Instruct-abliterated-GGUF
• EXL2: Apel-sin/llama-3.1-8B-abliterated-exl2

Use Cases

This model is particularly suited for applications where the default content moderation of standard instruction-tuned models is too restrictive, enabling more open-ended and less filtered text generation.