Datasets:
Data-quality audit on jaco_play — 695 REVIEW episodes out of 1,085
I ran a trajectory-only data-quality audit on this dataset (1,085 episodes) using RDA, an open-source CLI tool: https://github.com/liesliy/rda
Results:
- 390 PASS / 695 REVIEW / 0 EXCLUDE
- 11,958 action spikes across 77% of episodes
- Median idle: 74.1%
Integrity (NaN/Inf, timestamps, schema) is clean across all 1,085 episodes. This is the largest dataset in my benchmark, and it also has the highest absolute number of REVIEW episodes (695) and spike count (11,958).
Full benchmark: https://github.com/liesliy/rda/blob/main/docs/benchmark.md
Question: With 695 flagged episodes, has anyone tried selective training — training only on the 390 PASS episodes vs. all 1,085? Would anyone be willing to spot-check a sample of the REVIEW episodes to see whether the flags correspond to lower-quality demonstrations? That feedback would directly validate the tool's accuracy.