VerusCommunity/llama-3-verus-8-epochs-revision-1

Llama 3 Verus 8B: Specialized for the Verus Protocol

Llama 3 Verus 8b is an 8 billion parameter language model built upon the Meta Llama 3 architecture, specifically fine-tuned to provide comprehensive information about the Verus open-source multi-chain blockchain protocol. Developed by Evan Armstrong for the Verus Community, this model is designed to be a dedicated resource for understanding Verus.

Key Capabilities & Features

• Verus Protocol Expertise: Trained extensively on QA-style data generated by Verustoolkit, enabling it to memorize facts and answer questions about Verus.
• RAG Compatibility: Engineered to work effectively with Retrieval Augmented Generation (RAG) to enhance accuracy and mitigate hallucinations, especially at lower quantization levels.
• Standalone Performance: Capable of achieving good performance without RAG when sufficient latent space activation is provided via the system prompt.
• Open-Source: The entire development process, including data generation code, datasets, and training configurations, is open-source.
• Quantization Support: Includes Axolotl configurations and GGUF quantizations created using the imatrix method, with the imatrix itself also provided.
• ChatML Prompt Template: Utilizes the ChatML prompt template for conversations.

Recommended Usage

• Temperature: A low temperature, such as 0.05, is recommended for factual consistency.
• Quantization: For optimal precision and quality, the ggml-model-Q8_0.gguf quantized version is suggested for computer users.
• LM Studio: A specific LM Studio config is provided in the files section for ease of use.

This model is ideal for developers and users seeking in-depth, accurate information about the Verus blockchain protocol.