bs-vision-train-v03 / README.md
shaban2024's picture
v03 kartica (split po slici)
a79910c verified
|
Raw History Blame Contribute Delete
875 Bytes
metadata
language:
  - bs
license: cc-by-4.0
task_categories:
  - image-text-to-text
  - visual-question-answering
  - document-question-answering
tags:
  - bosnian
  - ocr
  - chart-qa
  - synthetic
  - bh-language-ai

bs-vision-train-v03 — sintetički bs trening set (2179 uzoraka, split po slici)

Isto kao v02 (iste slike, seed 43), ali split je PO SLICI pa val ne curi u train (v02 je imao 161 sliku u oba splita). Generator: scripts/make_synth_bs.py u https://github.com/BH-Language-AI/bh-vision.

  • data/synth3/train.jsonl (1962) + data/synth3/val.jsonl (217), slike data/synth3/img/ (1200 jpg, 32MB)
  • Tipovi: računi, fakture-tabele, obavještenja, izvodi (čćdžšđ, 30% rotacija); bar + line chartovi
  • Format: image-text-to-text chat JSONL. NIKAD miješati sa held-outom ni proxy evalom.
  • Za trening koristiti v03 (v01/v02 zastarjeli zbog curenja).