mymedieval-nli-clf

NLI-as-classification dataset derived from LULab/myMediEval_NLI.

• text: "Premise: {premise}\nHypothesis: {hypothesis}", both in Burmese.
• nli_label: one of Entailment, Contradiction, Neutral.

Built by reshaping the source dataset's wide format (one Burmese premise per row, with separately-authored Entailment/Contradiction/Neutral hypothesis columns) into long format: 300 premises x 3 relations = 900 rows, balanced 300/300/300. Only the Burmese premise column (my) was used; the parallel English (eng) text and annotator-specific columns (conv_ai, sentiment_ai, sentiment_min_thiha_tun, sentiment_thet_htet_san, conv_min_thiha_tun, speaker) from the source were dropped.

This dataset is under Non-Commercial (CC BY-NC-SA 4.0) License. No commercial use is allowed and only for Research and Development.

Known data-quality note

As of this snapshot, 2 of the 300 source premises have byte-identical hypothesis text reused across two different relation columns (one row has Entailment == Contradiction, another has Contradiction == Neutral), likely an upstream annotation/generation artifact. These 6 affected rows (2 premises x 3 relations) were kept in rather than silently dropped -- filter on duplicate text values across nli_label groups if you need a strictly clean split.