Problem

The data_path_prefix in dexbotic/data/data_source/*_official.py files uses singular video directory name, while the conversion scripts (convert_rlds_to_dexdata.py and
convert_lerobot_to_dexdata.py) output to plural videos/ directory. Additionally, the JSON output directory naming is inconsistent: conversion scripts output to jsonl/ but data_source configs
expected various directory names. Although dex_dataset.py uses recursive search to find .jsonl files, unified naming convention is still preferable for consistency. This inconsistency causes path
mismatch during training.

Root Cause

1. Conversion scripts output to videos/ but data_source configs expected video/
2. Conversion scripts output to jsonl/ but data_source configs expected various naming conventions

Technical Evidence

In dexbotic/data/dataset/transform/multimodal.py:123, video paths are resolved by joining data_path_prefix with frame['url']:

video_url = os.path.join(data_path_prefix, frame['url'])

The data_path_prefix is sourced from data_source/*_official.py configs and passed through meta_data. When frame['url'] contains relative paths like dataset_name/episode.mp4 (as generated by conversion
scripts), the resulting path depends entirely on data_path_prefix being correct.

If data_path_prefix is video but the actual directory is videos, the path resolution fails.