surindersinghssj/gurbani-kirtan-yt-captions-300h

---
configs:
- config_name: default
data_files:
- split: train
path: "data/*.parquet"
---

# Gurbani YouTube caption dataset

Built incrementally by `scripts/bulk_yt_caption_pipeline.py` using the
native-webm + UI-transcript-line caption-mode pipeline (see
`feedback_yt_caption_pipeline.md` in the agent memory for the method).

Each parquet shard in `data/` is an **additive batch**; new shards are
uploaded without rewriting previous ones. `datasets.load_dataset(repo)`
returns the union of all shards as a single `train` split thanks to the
glob in the YAML config above.

- sources: YouTube auto-captions (`pa-orig`) from per-video provenance
in the `video_id` column
- pipeline rules: native `webm/opus` download on IPv4; no resample at
master stage; clip boundaries = YouTube transcript-panel UI lines
(pair consecutive content events); `caption_offset_s = 0`
- Last snapshot: 2026-04-19 00:49:36