More validation data improves checkpoint choice in fine-tuning AI models
From Checkpoint Variation to Selection Gains in Supervised Fine-Tuning
Machine LearningArtificial Intelligence
Summary
When training AI models, many saved versions (checkpoints) are created, but only one is picked to keep. This paper studies how using more examples to test these checkpoints can help pick better ones. The researchers found that using more validation data improves selection and that some new ways of choosing checkpoints work better than just picking the one with lowest loss. However, it is still unclear if these better choices are always better than just keeping the final model at the end of training.
What this means in practice
- •For machine learning engineers: Use larger validation sets and generation-based rules to select better checkpoints during model fine-tuning for improved accuracy.
- •For ai product teams: Improve model update decisions by implementing advanced checkpoint selection criteria beyond final checkpoint retention.
Authors
Yupeng Chang, Wenxuan Zhang, Yuan Wu
Abstract
Checkpoint selection is a routine decision in supervised fine-tuning (SFT): training produces multiple checkpoints, but only one is retained. Yet fixed-budget comparisons do not by themselves distinguish three empirical claims: whether more validation data improve checkpoint selection, whether a selection rule outperforms validation-loss selection, and whether it improves over simply retaining the final checkpoint. We therefore treat checkpoint selection as a finite-information decision problem. Holding completed training trajectories, candidate checkpoints, and independent test items fixed, we vary the validation budget and separately measure improvement from additional validation data, gain over negative log-likelihood (NLL) selection, and gain over the final checkpoint. Across 60 mathematical SFT trajectories and 19 configurations, increasing the validation budget from 32 to 305-313 examples raises independent-test accuracy by 0.32 percentage points (pp) for generated-accuracy selection and 0.29 pp for checkpoint agreement, with 95% configuration-bootstrap CIs of [0.10, 0.56] and [0.11, 0.50], respectively. At the full validation budget, the two generation-based rules outperform matched NLL selection by 0.71 and 0.85 pp, respectively, while their gains over the final checkpoint remain unresolved. A cross-domain replication on 12 newly trained Commonsense trajectories shows the same qualitative separation: increasing the validation budget from 32 to 1,024 questions improves generated-accuracy and checkpoint-agreement selection by 0.87 and 0.27 pp, while gains over the final checkpoint again remain unresolved. Together, these results show that benefiting from more validation data, outperforming NLL selection, and outperforming the final checkpoint are distinct empirical claims that require separate evidence.