Post
A dangerous DK1 data-quality problem appears when batch export finishes without obvious errors but action counts, frame counts, or episode totals no longer agree across files. The dataset looks complete until downstream checks reveal inconsistent lengths.
How are you diagnosing DK1 batch exports where action and video counts do not match?
Please share how you audit exported sessions, which mismatches are still repairable versus discard-worthy, and what automated checks prevent count inconsistencies from reaching training or benchmarking.
If you reply, include one exact count-mismatch symptom and one exact export-audit check that exposed it.