EpisodeVault: Open-source tool to debug LeRobot model regressions

Rohan-Prabhakar released EpisodeVault, an open-source Python library designed to pinpoint why a LeRobot model's performance degraded after retraining. The tool addresses a common pain point: developers often know a model got worse but lack visibility into which specific dataset changes caused the regression. EpisodeVault provides four commands: `track` to initialize a dataset, `commit` to version it, `diff` to compare two versions with detailed metrics, and `blame` to trace a model back to its training dataset version.

In a real-world test on two LeRobot datasets, the diff command revealed that version 2.0 removed 7 episodes, with a 75% drop in kitchen_grasp episodes (4 to 1) and a success rate plunge from 0.88 to 0.38. The 'blame' command integrates with training scripts via `import episodevault as ev; ev.log_training_run(...)` and later outputs the dataset diff automatically. EpisodeVault is compatible with any local LeRobot v3 dataset and has been tested on four Hub datasets (e.g., aloha, so100) with parse times as low as 0.35s for 25 episodes. Install with `pip install episodevault` (Python 3.10+).