LOW-RESOURCE NLP / MULTILINGUAL DATA CURATION
Sambalpuri Language Resources
Supporting underrepresented languages through data and evaluation.
- Examples in the shared resource
- ~20,000
- Odia, Sambalpuri, and English
- 3
- Output comparison for model selection
- LLM
My contribution
I collected and organized data for shared low-resource NLP research on an approximately 20K-example Odia–Sambalpuri–English resource. I also compared LLM outputs to support model selection.
The work brings together multilingual data preparation and practical evaluation, helping make a regional language better represented in AI research.
Evaluation starts with the data
My focus was on organizing the language resource and reviewing model outputs. Clear examples, consistent language pairs, and careful comparisons help make model-selection decisions more useful.
The resource size describes the shared research dataset. It is not a model-performance score or an independently verified benchmark.
Related work: evidence-grounded AI
My work on Shivru AI and Advotac applies a similar concern for evidence: retrieved content and generated findings should stay connected to their source documents.