NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·1h ago
AI Research
arXiv CS.AI·May 26·5 min read
When Does Synthetic Patent Data Help? Volume-Fidelity Trade-offs in Low-Resource Multi-Label Classification
TL;DR
In low-resource multi-label classification, there is uncertainty about when synthetic patent data can enhance performance. The study utilized six open-source large language models (LLMs) to generate synthetic data for classifying 64 patent labels.
✦ Why It Matters
Engineers can optimize multi-label classification by strategically combining real and synthetic data to enhance performance.
Key Takeaways
- Use synthetic data to improve classification in low-resource settings.
- Combine 20-30% real data with 70-80% synthetic data for best results.
- Monitor the balance between data volume and fidelity for optimal performance.
- Experiment with different LLMs to find the best fit for your classification tasks.
How It Works
The study employs six LLMs, ranging from 3.8B to 12B parameters, to generate synthetic patent data. It tests both full-synthesis generation and paraphrasing methods, assessing their impact on classification performance across various data regimes.