Datasets:
Data-quality audit on svla_so101_pickplace — 90% REVIEW, 100% spike, 86.7% idle
I ran a trajectory-only data-quality audit on this dataset (50 episodes, SO-100/so100_follower) using RDA, an open-source CLI tool: https://github.com/liesliy/rda
Results:
- 5 PASS / 45 REVIEW / 0 EXCLUDE
- 260 action spikes across 100% of episodes
- Median idle: 86.7%
Integrity (NaN/Inf, timestamps, schema) is clean across all 50 episodes. But the behavior layer flags 90% for review — this dataset has the worst combination in my benchmark: 100% spike rate AND 86.7% idle.
Full benchmark (11 datasets, 4,909 episodes): https://github.com/liesliy/rda/blob/main/docs/benchmark.md
Context: If anyone is fine-tuning VLA models on community SO-100 data, this is what you're inheriting — high idle means the policy sees "do nothing" most of the time, and 100% spike rate means action discontinuities in every single episode. Has anyone noticed convergence issues when training on this dataset? Would you be willing to spot-check 5–10 of the REVIEW episodes to confirm whether the flags match your experience?