Reimagining service delivery in the agentic era with Google Public Sector
cloud.google.com·1d ago
AI Research
arXiv CS.AI·Jun 10·5 min read
Supervised Fine-tuning with Synthetic Rationale Data Hurts Real-World Disease Prediction
TL;DR
Real-world disease prediction models often struggle due to a lack of high-quality training data. Researchers explored the impact of supervised fine-tuning using synthetic rationale data, which is artificially generated explanations for model decisions.
✦ Why It Matters
Engineers should prioritize high-quality real-world data over synthetic alternatives for training disease prediction models.
Key Takeaways
- Avoid using synthetic rationale data for training clinical prediction models.
- Focus on label-only fine-tuning for better performance in disease prediction.
- Utilize rationales as inference-time demonstrations to enhance model accuracy.
- Investigate the structural conflicts in model training approaches.
How It Works
Supervised fine-tuning typically involves training models not only on what to predict but also on why those predictions are made, using synthetic rationale data. However, this study reveals that while the rationales are accurate, they create a conflict in the model's optimization process, leading to poorer performance.