TriMixGen-Indic @ FIRE 2026 - Subtask B: Word-Level Language Identification

FIRE 2026 track on identifying the language of each word in Romanized Hindi-Bengali-English and Hindi-Gujarati-English sentences. The evaluation scripts can be downloaded from https://github.com/Pruthwik/Evaluation-Scripts-For-TriMixGen-Indic-Shared-Task-FIRE-2026 repository. The scoring for subtask A uses multiple transformer models and multiple evaluation metrics. The submissions for subtask A will take upto 24 hours to get reflected on the codabench leaderboard. The scoring for subtask B uses the macro F1-score metric and the submissions will be reflected on the codabench leaderboard instantly.