README.md

metadata

license: cc-by-4.0
task_categories:
  - image-text-to-text
tags:
  - document-understanding
  - receipts
  - cord

CORD structured-extraction (LoRA training set)

200 train + 20 val examples derived from CORD (naver-clova-ix/cord-v2, train split). Each example pairs a preprocessed receipt image with the target extraction JSON (receipt fields + line items in the pipeline's schema).

• prompt.txt - the shared instruction (schema skeleton injected)
• train.jsonl / val.jsonl - lines of {"id", "image": "images/..png", "target": "<json>"}
• images/ - the preprocessed pages (deskew / resize<=1536 / denoise)

Used to QLoRA-fine-tune sarcasticcoder/qwen2.5-vl-7b-cord-lora.