YouTube comment dialogues: 399 parent comment – reply pairs from a socio-political discussion, translated into English

399 "parent comment – reply" pairs from the comments under one YouTube video on a socio-political topic, translated into English. The original comments are predominantly in Russian. Sampling. The pairs are an enriched subsample of a corpus of about 3,700 pairs. A random sample, distributed uniformly over the publication time of the comments, was enriched with manually selected extreme cases: clearly templated or coordinated replies and clearly live dialogue. Part of the average examples was then removed to keep the size. The row order is therefore not random with respect to how templated or live a reply is. Parent. For a top-level comment the parent is the description of the video. For a reply that does not start with an @-mention it is the root comment of the thread. For a reply that starts with an @-mention it is the comment of the mentioned user, found by walking back through the thread, or the root comment if that user is not in the thread. Translation. The texts were translated by the GPT6-Astra model under an instruction to render the meaning rather than the letter, including the deliberate spelling distortions, character substitutions and allusions with which the authors evade moderation. Anonymization. In line with the YouTube data policy, the original comments are not published. The file contains no user names, pseudonyms, comment identifiers, links, dates or like counts. Columns: - id: row number, 0–398; - parent_text_clean: the parent (see above); - textOriginal_clean: the reply; - _parent_method: top_level (112 rows), second_level (77) or mention_backtrack (210). The data set was used in the article "When Sorting Fails: Ranking from LLM Pairwise Comparisons Under High Tie Rates" (reference to be added), where large language model judges compared the relevance and authenticity of the replies.