Hugging Face Overview
YeungNLP/firefly-qwen1.5-en-7b is a 7.7 billion parameter language model derived from Qwen1.5-7B. It has been fine-tuned by YeungNLP using English instruction data to function as a helpful and harmless AI assistant. A notable aspect of its development is the efficient training process, utilizing QLoRA on a single V100 GPU.
Key Capabilities & Performance
This model demonstrates competitive performance, outperforming several established models like Qwen1.5-7B-Chat, Gemma-7B-it, and Zephyr-7B-Beta on the Open LLM Leaderboard. Specifically, firefly-qwen1.5-en-7b achieved an average score of 61.44 on the leaderboard, with strong results in MMLU (61.67) and GSM8K (55.34). While primarily trained with English data, it retains some proficiency in Chinese due to its Qwen1.5 base.
Training Details
The model underwent a supervised fine-tuning (SFT) stage, followed by a Direct Preference Optimization (DPO) stage for the firefly-qwen1.5-en-7b-dpo-v0.1 variant. Both stages were conducted with QLoRA on a single V100 GPU, highlighting an efficient training methodology. Key hyperparameters included a learning rate of 2e-4, a total batch size of 32, and a max sequence length of 2048 for SFT and 1600 for DPO.
Good for
- English Instruction Following: Excels at responding to English prompts as a helpful assistant.
- Resource-Efficient Deployment: Developed with efficient training methods, suggesting potential for optimized inference.
- Competitive General Performance: Offers strong benchmark scores compared to other 7B-class models.