NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·1h ago
AI Research
arXiv CS.AI·Jun 10·5 min read
Provenance-Grounded Gating and Adaptive Recovery in Synthetic Post-Training Data Curation
TL;DR
Existing synthetic data curation methods often overlook the importance of grounding filtering signals in source evidence and recovering rejected samples. This study introduces a method combining provenance-grounded gating and adaptive recovery strategies to enhance data curation.
✦ Why It Matters
Engineers can implement adaptive recovery techniques to improve data curation processes and enhance model performance.
Key Takeaways
- Incorporate source evidence in filtering processes.
- Utilize adaptive recovery strategies for rejected samples.
- Test various gate configurations for optimal results.
- Focus on generator scale for improved fine-tuning quality.
Full Summary
Synthetic data curation typically relies on filtering generated samples using reward models or large language model (LLM) judges, but it often fails to consider the source evidence behind each sample. This research investigates two key aspects: the grounding of filtering signals in source provenance and the systematic recovery of rejected samples.
The authors developed a controlled study that tested various gate configurations and recovery strategies using adversarially injected corpora to establish ground-truth failure labels. Findings indicate that grounding filtering in exact source provenance enhances the effectiveness of stronger judges, while different gates reject largely distinct sample populations, necessitating the use of both.
An adaptive recovery pipeline that combines failure diagnosis with targeted regeneration significantly outperformed naive resampling, achieving higher yield, recovery rates, and injection recall. Ultimately, the quality of downstream fine-tuning was primarily influenced by the scale of the generator, with filtration and recovery conditions playing a secondary role.