All projects
Generative AI & LLMs

JuktiLLM — RL for Bangla Reasoning

Ongoing

JuktiLLM: an ongoing curriculum-trained small language model for mathematical reasoning in Bangla, a low-resource morphologically rich language.

JuktiLLM — Bangla mathematical reasoning

The idea

Large language models reason well in English but fall sharply behind in Bangla — a language spoken by 230M+ people yet largely missing from the mathematical-reasoning literature. JuktiLLM (Bangla for “reasoning”) is an ongoing research project that trains a compact model to reason mathematically in Bangla through a staged curriculum, where each stage adds a progressively finer-grained learning signal. The goal is competitive Bangla math reasoning without frontier-scale model capacity.

Why it's hard

Bangla's agglutinative morphology makes correct reasoning traces structurally longer, which quietly confounds the reward signals standard reinforcement-learning methods rely on. Jukti's approach is designed to separate “was the answer correct” from “was it wordy,” so valid Bangla reasoning isn't penalised for length. (Method details are held back while the work is under submission.)

Tech stack & key skills

Core tools, methods and skills demonstrated in this project:

LLM post-trainingReinforcement learningCurriculum learningProcess reward modelsBangla NLPPyTorchQwen2.5-1.5B