The dataset provided in this competition consists of multilingual news articles labeled with topic categories. Each example corresponds to a single news item written in one of several African languages.

### Data Structure

Each row in the dataset contains:

1.  `sample_id`: A unique identifier for the article.
2.  `text`: The news article content or excerpt.
3.  `language`: The language code of the article.
4.  `label` (training only): The integer topic class identifier.
5.  `label_name` (training only): The human-readable topic name.

The test set contains all columns except the label fields. Participants must predict the integer topic label for each test example.

### Training Data

The labeled training data reflects a few-shot setting. The number of labeled samples per language is intentionally limited to encourage:

1.  Cross-lingual transfer learning
2.  Robust multilingual representations
3.  Careful handling of class imbalance

Participants may use the training data to train supervised models and may design additional validation splits for model selection.

### Test Data

The test set includes multiple languages and may exhibit distribution shift relative to the training data. Differences may include:

-  Variation in vocabulary and writing style
-  Differences in publisher characteristics
-  Changes in topic distribution across languages

Participants must submit predictions for all rows in the test set.

### Label Space

Each article belongs to exactly one topic category. Labels are represented as integer class IDs. The mapping between integer labels and category names is provided in the dataset.

### Public and Private Splits

The test set is internally divided into:

1.  A public subset used for leaderboard scoring during the competition.
1.  A private subset used for final ranking.

The split is defined in the solution file and is not visible to participants.

### Important Notes

-> Each `sample_id` appears exactly once in the test set.
-> Submissions must contain exactly one prediction per `sample_id`.
-> No external test labels are provided.

This structure ensures a fair evaluation of multilingual classification performance under realistic data constraints.