CALL FOR PARTICIPATION

BanglaKnight: Shared Task 1 at the Third Workshop on Bangla Language Processing (BLP 2027), co-located with NAACL 2027, San Francisco, June 1-5, 2027

Website: https://banglaknight.github.io/

Register on CodaBench: https://www.codabench.org/competitions/18411/

THE TASK

Safety classifiers that screen requests before an LLM answers are mostly trained on English. Bangla is often typed in Latin letters or mixed with English, Hindi or Urdu, and a harmful request can get past a filter when it is spelled differently.

BanglaKnight asks participants to label each Bangla prompt with one of four risk levels:

  0 Safe, 1 Offensive, 2 Harmful, 3 Severe

Every prompt appears in three written forms (Bangla script, romanized Bangla and code-mixed text) with the same label, and a system should give the same answer for all three. Systems are ranked by Defense Score: the mean macro-F1 over the three forms, multiplied by (1 - the share of unsafe prompts labeled Safe).

AVAILABLE NOW

- Training and development data

- A blind dev-test set for practice submissions on CodaBench

- A starter notebook with a BanglaBERT baseline that runs in about 8 minutes on a free Colab GPU

IMPORTANT DATES (all deadlines 11:59 PM AoE)

- Development phase: now until January 7, 2027

- Test phase: January 8-24, 2027

- Registration deadline: January 17, 2027

- Official results: February 3, 2027

- System papers (up to 4 pages) due: February 15, 2027

- Workshop: June 1-5, 2027, at NAACL 2027 in San Francisco

Individuals and teams of up to 5 are welcome. Please send questions to the CodaBench forum: https://www.codabench.org/forums/18099/

On behalf of the BLP 2027 organizers (https://blp-workshop.github.io/)