🇵🇭 FilBench - Can LLMs Understand and Generate Filipino?

As large language models (LLMs) become increasingly integrated into our lives, it becomes crucial to assess whether they reflect the nuances and capabilities of specific language communities. For example, Filipinos are among the most active ChatGPT users globally, ranking fourth in ChatGPT traffic (behind the United States, India, and Brazil [ 1 ] [ 2 ]), but despite this strong usage, we lack a clear understanding of how LLMs perform for their languages, such as Tagalog and Cebuano. Most of the existing evidence is anecdotal, such as screenshots of ChatGPT responding in Filipino as proof that it is fluent. What we need instead is a systematic evaluation of LLM capabilities in Philippine languages.

That’s why we developed FilBench: a comprehensive evaluation suite to assess the capabilities of LLMs for Tagalog, Filipino (the standardized form of Tagalog), and Cebuano, on fluency, linguistic and translation abilities, as well as specific cultural knowledge.

We used it to evaluate 20+ state-of-the-art LLMs on FilBench, providing a comprehensive assessment of their performance in Philippine languages:

In the past few years, we have seen an increase in region-specific LLMs that target Southeast Asian languages (SEA-specific), such as SEA-LION and SeaLLM. These are open-weight LLMs that you can freely download from HuggingFace. We find that SEA-specific LLMs are often the most parameter-efficient for our languages, achieving the highest FilBench scores compared to other models of their size. However, the best SEA-specific model is still outperformed by closed-source LLMs like GPT-4o.

The authors would like to thank Cohere Labs for providing credits through the Cohere Research Grant to run the Aya model series, and Together AI for additional computational credits for running several open models. We also acknowledge the Hugging Face team, particularly the OpenEvals team (Clémentine Fourrier and Nathan Habib) and Daniel van Strien, for their support in publishing this blog post.

The FilBench evaluation suite contains four major categories–Cultural Knowledge, Classical NLP, Reading Comprehension, and Generation–divided into 12 tasks. For example, the Classical NLP category includes tasks such as sentiment analysis, whereas Generation tasks include different aspects of translation. In order to ensure that these categories reflect the priorities and trends in NLP research and usage, we curate them based on a historical survey of NLP research on Philippine languages from 2006 to early 2024. (Most of these categories exclusively contain non-translated content to ensure faithfulness to the natural use of Philippine languages.)

Each of these categories provides an aggregated metric. To create a single representative score, we compute the weighted average based on the number of examples in each category, which we call the FilBench Score.

To simplify usage and set up, we built FilBench on top of Lighteval , an all-in-one framework for LLM evaluation. For language-specific evaluation, we first defined translation pairs from English to Tagalog (or Cebuano) for common terms used in evaluation such as "yes" ( oo ), "no" ( hindi ), and "true" ( totoo ) among others. Then, we used the provided templates to implement custom tasks for the capabilities we care about.