The Bengali NLP Safety Challenge 2026 introduces a shared task on Fine-Grained Bengali Toxicity Understanding. Online platforms increasingly rely on automated moderation systems to identify harmful, abusive, and unsafe content. Despite substantial progress in high-resource languages, toxicity understanding in Bengali remains challenging due to limited annotated resources, linguistic diversity, informal writing styles, implicit expressions, sarcasm, code-mixing, and culturally dependent forms of harmful language.
This challenge is designed to evaluate whether computational models can move beyond conventional binary toxicity detection toward a more detailed understanding of how harmful meaning is expressed. Rather than asking only whether a Bengali text is toxic, participating systems are expected to answer two complementary questions: Is the text toxic? And how is that toxicity expressed?
The competition consists of two closely related subtasks, evaluated together on a single Kaggle leaderboard. The central objective is to encourage systems that perform both coarse toxicity detection and fine-grained attribute understanding, aiming to develop robust, context-aware, and fine-grained Bengali language safety systems.
We welcome participation from researchers, students, engineers, practitioners, and NLP enthusiasts interested in Bengali NLP, online safety, responsible AI, and harmful-language understanding.
The first subtask evaluates whether a model can determine whether a given Bengali text contains toxic content.
* Input: Bengali Text
* Output: is_toxic (0 = Non-Toxic, 1 = Toxic)
This subtask evaluates the model's ability to distinguish harmful or abusive Bengali content from non-toxic language.
Toxicity is not expressed in only one way. A single Bengali comment may contain multiple toxic characteristics simultaneously. This subtask requires participants to identify one or more fine-grained toxicity attributes associated with each text.
The six prediction targets are:
* Implicit Toxicity (implicit_toxicity)
* Sarcasm (sarcasm)
* Profanity (contains_profanity)
* Threatening Content (threatening)
* Identity Attack (identity_attack)
* Sexual Content (sexual_content)
This is a multi-label classification task, meaning that several attributes may apply to the same text simultaneously. For example, one comment may contain both profanity and an identity attack, while another toxic comment may contain no explicit abusive words but still communicate harmful meaning through sarcasm or implicit toxicity.
Both subtasks contribute equally to a single combined leaderboard score.
Participants must submit predictions for a hidden-label Test Set containing 1,000 Bengali text samples. All predictions must be binary (0 or 1) for the seven target variables.