Hugging Face NeverSleep/Lumimaid-v0.2-70B is a 70 billion parameter instruction-tuned causal language model based on Meta-Llama-3.1-70B-Instruct, developed by NeverSleep. This iteration, Lumimaid 0.2, features a significantly refined dataset with extensive cleaning to remove 'sloppy' chat data. It is designed for general conversational tasks, leveraging a diverse collection of high-quality instruction datasets for improved performance.
Lumimaid-v0.2-70B Overview
Lumimaid-v0.2-70B is a 70 billion parameter instruction-tuned model developed by NeverSleep, building upon the Meta-Llama-3.1-70B-Instruct architecture. This version represents a substantial improvement over its predecessor, Lumimaid 0.1, primarily due to a rigorous dataset refinement process. The developers, Undi and IkariDev, focused on meticulously cleaning and curating the training data to eliminate low-quality or 'sloppy' chat examples, aiming for a higher standard of conversational output.
Key Capabilities
- Enhanced Instruction Following: Benefits from a cleaned and expanded dataset, leading to more precise and coherent responses to instructions.
- Llama-3-Instruct Prompt Template: Utilizes the standard Llama-3-Instruct prompt format for consistent interaction.
- Diverse Training Data: Trained on a wide array of datasets including Gnosis, Luminous_Opus, Synthetic-Dark-RP, Synthetic-RP, Sonnet3.5-SlimOrcaDedupCleaned, Opus-WritingPrompts, and various other instruction and chat datasets, ensuring broad knowledge and conversational ability.
Good For
- General-purpose conversational AI applications requiring high-quality, instruction-tuned responses.
- Scenarios where clean and refined chat data is crucial for model performance.
- Developers seeking a robust Llama-3.1-based model with improved data quality for fine-tuning or direct deployment.