Datasets:
Update take captions with Claude-generated task/subtask descriptions
#3
by yukangcao - opened
Regenerated every take's caption/llm_description*.json from timeline/videos/homie_front_left.mp4 using Claude (claude-fable-5-1, and claude-opus-5 for takes 000233/000234/000461-000467).
- 628 caption files across 467 takes; same 'Task: ... / Subtasks: N. [start-end] ...' format.
- Clips over 200 s are split into parts with 10 s overlap (llm_description_partN.json); subtask times are relative to each part's start, and start/end give the part's position in the clip.
- Removes the duplicated llm_description_part00N.json series.
- take-000384 is unchanged: it has no timeline/videos/homie_front_left.mp4.
acerobotics2025 changed pull request status to merged