TeluguLLMResearch/MATA

MATA (మాట): Mindful Assessment of the Telugu Abilities of Large Language Models

• A new evaluation dataset specifically designed to assess a diverse spectrum of linguistic competencies in Telugu.
• A test suite that directly engages with Telugu-specific grammatical structures, semantic patterns, and culturally embedded language use by sourcing our questions from school grade language textbooks and higher level exams.
• 729 carefully curated multiple-choice and open-ended questions.
  • Approximately 25% are multiple-choice questions and the rest are open-ended.
  • A range of difficulty from lower grade (simpler syntax and vocabulary) through high-school texts (advanced grammar and semantics).

Citation:

@inproceedings{kranti-vajjala-2026-mata,
    title = "MATA (మాట): Mindful Assessment of the Telugu Abilities of Large Language Models",
    author = "Kranti, Chalamalasetti  and
      Vajjala, Sowmya",
    booktitle = "Proceedings of the 2026 International Conference on Language Resources and Evaluation (LREC 2026)",
    month = may,
    year = "2026",
    address = "Palma, Mallorca (Spain)",
    url = "https://arxiv.org/abs/2508.13526",
}