[TRLC-DK1] Episode boundaries and dataset QA for builders labs (intermediate)
TRLC-DK1 discussion on episode boundaries, bad-demonstration filtering, dataset QA, and practical checks before uploading or training.
DK1 data collection gets messy when teams are not consistent about where an episode begins, where it ends, and which runs should be excluded before training.
How are you defining episode boundaries and filtering low-quality demonstrations in your DK1 workflow?
Please share practical checks for resets, idle frames, partial failures, operator hesitation, and what your QA pass looks like before a dataset is considered usable.
If you reply, include one exact QA rule that caught a bad run and one rule that turned out to be too strict.
A lot of searchers land on this question because they already have data but do not trust it. Include how you make that decision.
If your team uses a manual review sheet or simple heuristics before training, post that process. It is often more useful than a perfect theory answer.
Good replies will distinguish between a failed task that is still informative and a recording that is simply too broken to keep.