Reproduce “AlpaGasus: Training A Better Alpaca with Fewer Data” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Lichang Chen, Shiyang Li, Jun Yan, Hai Wang, Kalpa Gunaratna, Vikas Yadav, Zheng Tang, Vijay Srinivasan and 3 more
An automated data selection strategy, ALPAGASUS, leverages strong LLMs to filter low-quality examples from instruction fine-tuning datasets, leading to models that surpass the original Alpaca model while using only 17.75% of the data. This approach reduces training time by 5.7x and lowers costs, making LLM fine-tuning more efficient.
Key insights
- Data quality, rather than sheer quantity, is the dominant factor in achieving high performance during instruction fine-tuning of LLMs.
- Automated data filtering using powerful LLMs as judges provides an efficient and scalable alternative to manual curation for improving IFT dataset quality.
- Reducing the training data size through intelligent filtering significantly enhances training efficiency, reducing computational cost and time without compromising performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.