bs-vision-train-v01 / README.md
shaban2024's picture
v01 kartica
a4df88e verified
|
Raw History Blame Contribute Delete
1.1 kB
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-v01 — sintetički bs trening set (400 uzoraka)

Vlastiti generisani sadržaj (100% čista licenca, CC-BY-4.0) za prvi BH-Vision trening. Generator: BH-LANGUAGE-AI/bh-vision-training/scripts/make_synth_bs.py (seed 42).

  • data/synth/train.jsonl (360) + data/synth/val.jsonl (40), slike data/synth/img/ (300 jpg, 8MB)
  • Tipovi: računi/obavještenja sa čćdžšđ (OCR + prepisivanje) i bar-chartovi sa mjesecima (chart QA)
  • Format: image-text-to-text chat JSONL — messages[{role, content:[{type:image},{type:text}]}] + images (putanje relativne na root frameworka)
  • Font: DejaVu Sans (Bitstream Vera licenca); chartovi: vlastiti podaci, matplotlib
  • NIKAD miješati sa held-outom bs-vision-ocr-heldout-v01 niti proxy evalom (XM3600-hr, postocr-sr)
  • Trening: bh-vision-training/configs/smo*-qlora.yaml, start od najmanjeg (SmolVLM2-2.2B)