The dataset is currently empty. Upload or create new data files. Then, you will be able to explore them in the Dataset Viewer.

BH-Vision banner

Lang: bs Status: placeholder

bs-vision-ocr-heldout-v01 — STATUS: PLACEHOLDER, NEMA PODATAKA

Bosanski vision held-out eval set (dokumenti, računi, tabele, chartovi, rukopis sa čćdžšđ). NIKAD u train.

  • Verzija: v01 (frozen split kad se napuni; do tada prazan placeholder)
  • Format (plan): image-text-to-text chat JSONL — messages[{role, content:[{type:image},{type:text}]}] + images
  • Proxy eval do tada: google/xm3600 (hr), postocr-sr (latinica), tehnika na PubLayNet/PlotQA
  • Baze za trening: SmolVLM2-500M / 2.2B (start), Idefics3-8B (tek nakon pravne provjere), Pixtral-12B (opcionalno)
  • Odluke: BH-LANGUAGE-AI/reports/research/2026-11-11-vision-models.md, 2026-11-11-vision-datasets.md
  • Licenca cilj: CC-BY-4.0 za vlastite uzorke; tuđe slike samo uz provjeru OBA sloja prava (anotacije + slike)
Downloads last month
32