Firefly-Gemma-7B: A High-Performing Gemma-Based Assistant
YeungNLP/firefly-gemma-7b is an instruction-tuned large language model built upon Google's Gemma-7B. This model has been optimized using the Firefly training framework with QLoRA on a single V100 GPU, aiming to serve as a helpful and harmless AI assistant.
Key Capabilities & Performance
- Enhanced Performance: Firefly-Gemma-7B significantly outperforms several established models in its class, including the official
gemma-7b-it, zephyr-7b-gemma-v0.1, Qwen1.5-7B-Chat, and Zephyr-7B-Beta, as evidenced by its strong showing on the Open LLM Leaderboard. - Benchmark Scores: Achieves an average score of 62.93 on the Open LLM Leaderboard, with notable scores such as 79.77 on HellaSwag, 61.57 on MMLU, and 49.28 on GSM8K, indicating robust general reasoning and language understanding capabilities.
- Efficient Training: The model was efficiently trained using QLoRA, demonstrating that competitive performance can be achieved with accessible hardware.
Ideal Use Cases
- General AI Assistant: Excels in conversational AI scenarios, providing helpful and harmless responses.
- Benchmarking & Research: Its strong performance on public leaderboards makes it a valuable candidate for research and comparative analysis of instruction-tuned models.
- Resource-Efficient Deployment: Trained with QLoRA, it offers a potentially more efficient option for deployment compared to models requiring more extensive resources.