Batayan: A Filipino NLP benchmark for evaluating Large Language Models

Recent advances in large language models (LLMs) have demonstrated remarkable capabilities on widely benchmarked high-resource languages. However, linguistic nuances of under-resourced languages remain unexplored. We introduce BATAYAN, a holistic Filipino benchmark that systematically evaluates LLMs across three key natural language processing (NLP) competencies: understanding, reasoning, and generation. BATAYAN consolidates eight tasks, three of which have not existed prior for Filipino corpora, covering both Tagalog and code-switched Taglish utterances. Our rigorous, native-speaker-driven adaptation and validation processes ensures fluency and authenticity to the complex morphological and syntactic structures of Filipino, alleviating the pervasive translationese bias in existing Filipino corpora. We report empirical results on a variety of open-source and commercial LLMs, highlighting significant performance gaps that signal the under-representation of Filipino in pre-training corpora, the unique hurdles in modeling Filipino's rich morphology and construction, and the importance of explicit Filipino language support. Moreover, we discuss the practical challenges encountered in dataset construction and propose principled solutions for building culturally and linguistically-faithful resources in under-represented languages. We also provide a public evaluation suite as a clear foundation for iterative, community-driven progress in Filipino NLP. Comp. Task Dataset # test samples Target 1 Language Source 2 Adaptation Our contribution NLU PI PAWS (Zhang et al., 2019) 2,000 label (2) English English-sourced Native translation QA Belebele (Bandarkar et al., 2024) 900 span (4) Filipino English-adapted Native re-translation SA PH Elections (Cabasag et al., 2019) 5,160 label (3) Taglish Natively-sourced No adaptation TD PH Elections (Cabasag et al., 2019) 5,160 label (2) Taglish Natively-sourced No adaptation NLR CR Balanced COPA (Kavumba et al., 2019) 500 label (2) English English-sourced Native translation NLI XNLI (Conneau et al., 2018) 5,010 label (3) English English-sourced Native translation NLG AS XL-Sum (Hasan et al., 2021) 11,535 summary English English-sourced Native translation MT FLORES 200 (NLLB Team et al., 2022) 1,012 translation Filipino English-adapted Native re-translation 1 Number of options are shown in parenthesis 2 English-adapted: previously translated from English; English-sourced: originally in English; Natively-sourced: originally in Taglish