This competition addresses multilingual news topic classification in a few-shot setting. Each example consists of a news article (or excerpt) written in one of several African languages. The task is to assign the correct topic label to each article from a predefined set of categories.

The training data includes labeled examples across multiple languages, but the number of labeled samples per language is intentionally limited. This setup reflects realistic low-resource conditions where annotated data is scarce. Participants must therefore design models that can transfer knowledge effectively across languages and remain robust under data constraints.

The evaluation data contains examples drawn from diverse domains and writing styles. Some test examples may differ from the training distribution in terms of publisher characteristics, vocabulary usage, or stylistic variation. As a result, models that overfit to language-specific or domain-specific artifacts are unlikely to perform well.

Key challenges in this competition include:

1.  Learning shared representations across multiple languages.
2. Handling class imbalance within and across languages.
3. Generalizing under distribution shift.
4. Maintaining strong worst-language performance rather than optimizing for dominant languages only.

The problem is structured to encourage thoughtful modeling choices. Approaches such as multilingual transformer fine-tuning, contrastive learning, semi-supervised learning, self-training, domain adaptation, data augmentation, or language-aware calibration may provide meaningful improvements over the baseline.

Participants should pay close attention to validation strategy. Because performance is measured separately per language before aggregation, it is important to monitor per-language F1 scores during development. A model that achieves strong overall accuracy but collapses on one language will receive a reduced final score.

The provided baseline demonstrates the correct preprocessing pipeline, model training loop, and submission format. However, it is intentionally simple and leaves significant room for improvement. The competition rewards solutions that achieve both high average performance and balanced multilingual robustness.