mT5-base: topic classification + headline generation (Hausa, Igbo, Yoruba, Nigerian Pidgin)
One multi-task google/mt5-base model fine-tuned for the DSN Bootcamp Hackathon 2026 (LLM/Agent Track).
Code, results and write-up: https://github.com/faith-amanze/dsn-multilingual-topic-headline
Prompts
topic <lang>: <article text>headline <lang>: <article text>
<lang> is one of hau, yor, ibo, pcm. Max input 320 tokens, max target 40 tokens.
Results (dev set, 12% stratified hold-out of train)
Final 0.6919 (MacroF1 0.8709, GenScore 0.5129) vs 0.2520 for a majority-label / first-12-words baseline. The best checkpoint was selected on this same dev set, so scores are slightly optimistic.
Limitations
For topic prompts, the competition pipeline forced the output onto the 7 labels (business, health, politics, religion, sports, entertainment, technology); raw generate() can return other text. Headlines are capped at 40 tokens, so many end mid-sentence.