Ubvarol/watson-nli-xlm-roberta

🕵️ Contradictory, My Dear Watson — an NLI ladder

Sentence-pair natural language inference across 15 languages: given a premise and a hypothesis, predict entailment, neutral, or contradiction (Kaggle Contradictory, My Dear Watson, metric: accuracy).

The story is a three-rung ladder — multilingual DistilBERT 0.606 → XLM-RoBERTa base 0.691 → plus an XNLI warm-up 0.762 — and one finding worth keeping: the model scored 0.749 before seeing a single competition pair. Learning the task beat learning the dataset.

A static write-up; the notebook lives on GitHub.