Datasets:
Kirtan + Sung Bani: rajikaurrr channel, 15 videos (14 Shabad kirtan + 1 Salok Mahalla 9), whole channel in practice. Merge after #15
Owner's second data source: YouTube channel rajikaurrr (https://www.youtube.com/@rajikaurrr; 544 videos, slides with Gurmukhi + English translation). 15 diverse videos, one clip each: 14 Shabad kirtan in 1-kirtan-shabad (several with parmaan Shabads as their own shabads.csv rows) and Salok Mahalla 9 (20 min) in 3-bani-sung-like-kirtan/saloks-sung (bani:30).
Split: the singers are not named and no singer is visible in the frames, so the whole channel is in ONE split (dev, practice), keeping any singer out of both practice and test.
Keys: slide change (frame difference) + OCR of the yellow English row (tesseract on GCP after yellow-text isolation; no downloads or OCR on the Mac), matched to the DB English translations over the whole STTM DB, then inside the clip's Shabads; lines_source broadcast-caption. Independent verifier: 151/151 checked rows right (3 rows in YI-zFibFcSw may start up to ~3 s late where the publisher split one DB pangti over two slides). Runner apaath:. Built on PR #15: merge #11, #13, #14, #15 first. Existing manifest rows byte-identical. Tools: research/cloud-gab/hf-sung-bani/ (channel-ingest.sh, vm-fetch-ocr.sh, kmatch.js KMATCH_MODE=en).
Superseded by #19 (one combined PR on the current main that merges cleanly): https://huggingface.co/datasets/khalisfoundation/GurbaniAutopilotBench/discussions/19