jjjardev/tagasenti

TagaSenti is a dataset and AI model pair built for understanding sentiment in Tagalog and Taglish, the mix of Tagalog and English that many Filipinos speak and type online. The project targets a specific gap: most sentiment analysis tools are built for English text, and Tagalog or Taglish text often confuses them. At its core is a dataset of 35,686 sentences pulled from five different domains: e-commerce reviews, news, social media, adversarial examples designed to trip up a model, and synthetic sentences. Each sentence is labeled as negative, neutral, or positive. Over 9,300 rows, nearly a third of the dataset, were built specifically to test tricky edge cases across four generations of adversarial examples, meant to make the resulting model more reliable on real-world text rather than just easy cases. The dataset is hosted on Hugging Face and released under a CC BY-SA 4.0 license. Paired with the dataset is a fine-tuned model based on xlm-roberta-large, a 355 million parameter multilingual language model. On its own test set it reaches 86.6 percent accuracy with a macro-F1 score of 0.866. The team also tested it on a related language, Hiligaynon, without any additional training, and it still scored 62.4 percent accuracy, suggesting the model picks up patterns that carry over between related Philippine languages. The model weights are released under the Apache 2.0 license. The repository includes a full training script that loads the dataset directly from Hugging Face, splits it into training, validation, and test sets, cleans and normalizes the text (handling things like repeated letters and slang), tokenizes it, and trains the model with techniques like label smoothing and a cosine learning rate schedule. Running it requires Python 3.10 or newer and a GPU, since the original model was trained on a Tesla T4. The code itself is released under the MIT license, so the three parts of the project, data, model, and code, each carry their own license.