lenalibon/Meta-Llama-3-8B-Instruct-heretic-mlabonne
The lenalibon/Meta-Llama-3-8B-Instruct-heretic-mlabonne model is an 8 billion parameter instruction-tuned variant of Meta's Llama 3, specifically modified using the Heretic v1.2.0 tool to be a decensored version. It retains the 8192-token context length and optimized transformer architecture of the original Llama 3-8B-Instruct. This model is primarily intended for dialogue use cases where reduced refusal rates and broader response generation are desired, making it suitable for applications requiring less restrictive content filtering than the base Meta Llama 3-8B-Instruct.
Loading preview...
lenalibon/Meta-Llama-3-8B-Instruct-heretic-mlabonne Overview
This model is a decensored version of Meta's 8 billion parameter Llama 3-8B-Instruct, created using the Heretic v1.2.0 tool. It maintains the core architecture and 8192-token context length of the original Llama 3-8B-Instruct, which was developed by Meta and optimized for dialogue use cases through supervised fine-tuning (SFT) and reinforcement learning with human feedback (RLHF).
Key Capabilities
- Reduced Refusals: Significantly lowers the refusal rate compared to the original Meta-Llama-3-8B-Instruct (6/100 vs. 98/100), offering broader response generation.
- Instruction-Tuned: Optimized for assistant-like chat and dialogue applications.
- Strong General Performance: Inherits the strong benchmark performance of Llama 3-8B-Instruct across various categories, including MMLU (68.4), HumanEval (62.2), and GSM-8K (79.6).
- Text and Code Generation: Capable of generating both text and code outputs.
Good For
- Unfiltered Dialogue Applications: Ideal for use cases where a less restrictive content policy is preferred or required, such as creative writing, role-playing, or research into model biases.
- Assistant-like Chatbots: Suitable for building conversational agents that require high helpfulness and a lower propensity to refuse prompts.
- English-language Commercial and Research Use: Intended for applications primarily in English, with potential for fine-tuning for other languages under the Llama 3 Community License.
- Developers Seeking Llama 3 Performance with Modified Safety Alignment: Offers the robust capabilities of Llama 3-8B-Instruct with a modified safety profile.