HN Tube field note
How the HackerTube list was built
The source filters, AI stance labels and recurring-month recommendation score behind the public HackerTube dataset.
The public HackerTube export documents a multi-stage pipeline. It begins with the open-index Hacker News archive, identifies YouTube links surrounded by meaningful comment text, and keeps videos mentioned by at least three different people across at least three different months.
Mentions are labeled by stance
Retained comments were classified as strong recommendations, recommendations, neutral references, jokes or criticism. These labels were generated by a language model and can be wrong. HN Tube therefore exposes the source context and publishes only records with at least one positive label during its first phase.
The score rewards recurrence
The documented score is the number of distinct recommendation months plus half the number of distinct months containing a strong recommendation. It measures recurring positive attention, not video quality, factual accuracy or community-wide approval.
Known limits
At most twenty mentions per video are included in the export. Newer videos have had less time to accumulate recurring recommendations. Separate uploads can contain the same talk, and topic assignments and summaries were also generated by a model. Read the HN Tube methodology for our additional publishing and quality gates.