This is a non-destructive LeRobot v3.0 merge of six 200-episode benchmark
datasets. It contains 1,200 training episodes and 609,553 frames. Every
episode's task string is the assigned prompt from meta/benchmark/prompt_manifest_v3.csv.
The Coke and Pepsi objects are canonically labeled as a crushed Coke can and a
crushed Pepsi can. All rows are part of the training split.
Prompt manifest v3 applies the exact 1… See the full description on the dataset page:
https://huggingface.co/datasets/justintiensmith/VLA_Benchmark_Prompted_1200.
