Most language models are fluent in English and shaky in almost everything else. This competition puts that gap in front of you: you'll work with news articles in Hausa, Igbo, Yoruba, and Nigerian Pidgin — four of Nigeria's largest languages, and four that barely register in mainstream training data — and try to build something that genuinely works on them.
Each article needs two things done to it: pick the topic, and write a headline. Both are scored, across all four languages at once, and merged into a single number that determines your rank. Being excellent at labelling and useless at writing will not place you.
There is also a hard ceiling on firepower. Your scored model must be under 1 billion parameters, and everything — training, fine-tuning, inference — must run inside a single notebook on a free GPU (1× T4 on Colab, or 2× T4 on Kaggle). No clusters, no A100s, no checkpoints trained elsewhere. The constraint is the interesting part: the question isn't what a frontier model can do with these languages, it's what you can do with a small one and a free afternoon of GPU time.
Build a single sub-1B-parameter model that, given a Nigerian-language news article, both classifies its topic into one of 7 categories and generates a headline in the article's own language. Your rank is set by one merged score — half topic classification (Macro-F1), half headline generation (ROUGE-L and BERTScore blended) — computed across Hausa, Igbo, Yoruba, and Nigerian Pidgin together.
The competition data is derived from MasakhaNEWS (Adelani et al., 2023), a human-annotated news corpus covering 16 African languages. We have taken the four Nigerian languages, cleaned and shuffled them, and combined them into single files.
Three files are provided:
* train.csv and dev.csv: Contain five columns (id, language, text, headline, label).
* test.csv: Contains only id, language, and text. The headline and label values are withheld.
Note the language column is given to you at inference time. There is no language-identification step in this competition — you always know which language you are reading.
Task A — Topic classification
Given text, predict exactly one of seven labels:
* business
* health
* politics
* religion
* sports
* entertainment
* technology
Scored with macro-averaged F1 over the pooled test set. Macro-averaging is deliberate to ensure the model gets less frequent labels like technology right, rather than just guessing the most common label.
Task B — Headline generation
Given the same text, generate a headline in the article's own language. Scored with a blend of lexical overlap (ROUGE-L) and semantic similarity (BERTScore using a multilingual backbone).
Your leaderboard position is determined by a single number combining both tasks equally: * Macro-F1: For topic classification. * GenScore: A blend of ROUGE-L and BERTScore for headline generation.
Both components lie between 0 and 1, ensuring neither task dominates through scale alone. Ties break on Macro-F1 first, then smaller parameter count, then shorter runtime.
Submit one CSV file with two columns (id, prediction) in long format. Each test row produces two prediction rows, distinguished by an _topic or _headline suffix on the id.
Example:
id,prediction
hau_0001_topic,politics
hau_0001_headline,Gwamnati ta kaddamar da shirin tallafi ga kanana sana'o'i
ibo_0002_topic,sports
ibo_0002_headline,Nna ndi egwuregwu Naijiria meriri asompi ahu
pcm_0003_topic,health
pcm_0003_headline,Government open new hospital for Lagos
Topic predictions must be lowercase and drawn from the seven labels. Headline predictions are free text; quote any field containing a comma and never submit a blank.