Current Projects

Paper draft

Data scaling in bilingual machine translation

How does translation quality change as you add more parallel data in Nepali-English pair, and does that pattern hold across language pairs?

  • Extended the study to Hindi–English, a well-resourced pair, and Tamang–English, a truly underrepresented pair
  • Machine Translation
  • Scaling Laws
  • Low-resource
Ongoing

Adapting Gemma 4 to Nepali

Making open LLMs work better for low-resource languages, starting with Nepali.

  • Continued pre-training on ~1.76B Nepali tokens from many sources
  • ~500M English tokens replayed to limit catastrophic forgetting
  • Post-training (SFT and alignment) on top of the adapted model
  • LLMs
  • Continued Pre-training
  • SFT
  • Nepali
Paper draft

Nepali–English code-mixing

A new Nepali–English code-mixed dataset, with analysis and a benchmark.

  • Analysis of how code-mixing varies across different profession groups
  • Benchmark of current models on the data
  • Code-mixing
  • Dataset
  • Benchmark

What I'm Reading

  • Continued pre-training (CPT): data mixtures, replay strategies, and vocabulary or tokenizer adaptation for new languages
  • Supervised fine-tuning (SFT): instruction data for low-resource languages and how it interacts with CPT
  • Safety and guardrails for LLMs: how alignment and safety behaviour carry over, or fail to carry over, once a model is adapted to a new language

Open To

Graduate opportunities, research collaborations, and engineering roles in language technology.