Checklist BdAIO 2025 National Round · ML task
Constrained Headline Generator
Generate Bengali news headlines with a provided model while avoiding per-article banned words.
The task
Contestants receive a pretrained Bengali headline-generation model (bdaio-org/newspaper-title-titulm3b) and must run inference so that each generated headline is fluent and meaningful but contains none of the restricted words given for that article, even though those words are closely related to the content. No training is required.
The page illustrates this with a Bengali article about a new metro-rail route and the restricted words 'উদ্বোধন', 'ঢাকা', 'মেট্রোরেল', contrasting an acceptable headline with one that uses the banned words, and gives a script that loads the model in 8-bit and generates text.
Abridged by SOTA from the official materials. The official statement has the exact rules, and it wins wherever this summary differs.
At a glance
- You get
test.csv(news article content and a list of restricted words per article).- You submit
- A CSV with one generated headline per article (format as in
sample_submission.csv). - Scoring
- Mean over samples of (BERTScore F1 against the reference headline − penalty), where penalty = number of restricted words in the prediction / number of tokens in the prediction (custom metric 'modified bertscore').
- Rules
- Individual; up to 20 submissions per day.
- BdAIO rules (2026 page): all modelling on Kaggle; LLM assistants such as ChatGPT, Gemini or Claude are prohibited.
- Format
- BdAIO 2025 National Round, Kaggle competition 17 May 2025, 03:30–08:30 UTC (5 hours).