apify-youtube-transcripts-llm-training-datalisted

Build LLM training data from YouTube transcripts in bulk with the Apify Actor johnvc/YoutubeTranscripts. Feed it an array of video URLs and get one clean row per video, with non_timestamped transcript text ready for tokenization, timestamped snippets, language_code, and provenance metadata (title, channel_name, upload_date, view_count) for dataset documentation. Filter by language or translate every transcript into one target language for a consistent corpus. Use when the user wants LLM training data from videos, a fine-tuning or RAG corpus from YouTube, bulk YouTube transcripts for machine learning, or a text dataset from spoken video content without running speech-to-text. Pay-per-video billing, MCP-ready for Claude and other AI agents.
johnisanerd/claude-skill-youtube-transcripts-llm-training-data · ★ 0 · AI & Automation · score 66

Install: claude install-skill johnisanerd/claude-skill-youtube-transcripts-llm-training-data

# YouTube Transcripts as LLM Training Data Turn a list of YouTube videos into a clean text corpus for LLM fine-tuning, RAG, or analysis. One run takes an array of URLs and returns one row per video: the full transcript as a single text field, plus the provenance metadata (title, channel, upload date, views) a documented dataset needs. ## When to use this skill - The user wants LLM training data, a fine-tuning corpus, or RAG documents sourced from YouTube videos. - They want bulk transcripts from a channel dump, a playlist export, or a keyword-collected URL list. - They want spoken-video content as text without running their own speech-to-text. - They need a single-language corpus from mixed-language sources (translate everything to one target). Not for: discovering which videos to include (collect URLs first; see the chain table in `references/actor-index.md`), videos without caption tracks (those return error rows), or verbatim audio ground truth (captions are YouTube's own tracks, not fresh ASR). ## What you get (one row per video) The corpus fields: `non_timestamped` (the full transcript as one string, the training-text column), `language_code`, `source_type` (Manual or Auto-generated, a quality signal), `snippet_count`, `total_seconds`. The provenance fields for dataset documentation: `title`, `channel_name`, `channel_url`, `upload_date`, `view_count`, `like_count`, `tags`, `categories`, `url`, `video_id`. Optional alignment data: `timestamped` (snippets with `text`