Model Overview

umd-zhou-lab/recycled-wizardlm-7b-v1.0 is a 7 billion parameter auto-regressive language model developed by the UMD Tianyi Zhou Lab. It is built upon the Llama-2-7b architecture and fine-tuned using a unique "data recycling" approach with WizardLM (70k) data. This method, detailed in the paper "Reflection-Tuning: Data Recycling Improves LLM Instruction-Tuning," aims to enhance instruction-tuning efficiency and model performance.

Key Capabilities & Performance

This model showcases improved performance metrics compared to its base Llama-2-7b and the original WizardLM 7B model. Key benchmarks include:

• AlpacaEval: Achieved 78.88, significantly higher than WizardLM 7B's 67.64.
• Average Score (Open LLM Leaderboard): 56.21, an improvement over WizardLM 7B's 54.18.
• MMLU: Scored 48.35, surpassing WizardLM 7B's 42.70.

These results suggest that the data recycling methodology effectively boosts the model's instruction-following and general reasoning capabilities.

Intended Use Cases

This model is primarily intended for:

• Research: Exploring advanced instruction-tuning techniques and data efficiency in large language models.
• Chatbot Development: Serving as a foundation for experimental chatbot applications.
• Hobbyist Projects: Providing a capable base for natural language processing and AI enthusiasts.