01
Open weights
Open base models where possible, hosted inference where it makes sense, with an eye on cost per token in a peso market.
WHY
THE WAY FILIPINOS WRITE
Frontier models handle English fine and Tagalog badly. Bisaya, Cebuano and code-switched Taglish fall off a cliff. BudotsGPT and FilipinoGPT are our LoRA finetunes of open models for the languages and the register people use every day, in news, in tourism and online.
BUILD
HOW IT IS BUILT
02
LoRA finetuning
Filipino-context adaptation: instruction tuning with LoRA on local corpora.
03
Synthetic data
Generated and human-audited datasets covering Cebuano, Tagalog and Taglish register, plus domain sets for news and tourism.
04
Retrieval
Retrieval and summarisation layers feeding Balita.ph and internal newsroom tools.
TEST
HOW WE TEST IT
Benchmarks that test for local idiom, not just BLEU, and human raters from the region the text is for. A model that is fluent in Manila and lost in Cebu has not passed.
LoRAFinetuning
OSSBase model policy
PHHuman raters
MORE
ALSO IN GEN AI
- GenAI modelsWhat we ship: finetuning, synthetic data, AI in production pipelines
- AI & 3D worldsWorld models, 3D from a photo, digital humans
- NewsPHThe Balita.ph news platform
- The studio on GitHubgithub.com/budotsmediaph