working on low-resource language technologies that consider the context of the community and language >>> arxiv.org/abs/2610.25440
AI Weekly's analysis
→
- Fine-tuning a sub-3B vision-language model on 35 transcribed pages dropped bilingual Lakota archival OCR character error rate from 17.1% to 3.9%.
- Training on just nine pages added only 1.1 percentage points of CER over the full 35-page run, hinting the data cliff is shallow.
- Switching from gold to realistic automatic segmentation nearly doubled the CER of the best models, so idealized benchmarks overstate deployed accuracy.
Read full analysis →
