sroie-layoutlmv3

microsoft/layoutlmv3-base fine-tuned for key-information extraction on the SROIE 2019 dataset (Malaysian receipts). Extracts company, date, address, and total via BIO token classification.

• Base model: microsoft/layoutlmv3-base (~125M parameters)
• Task: token classification, 9 BIO labels (O + B-/I- for company/date/address/total)
• Training: 563 train / 63 validation SROIE receipts. AdamW, lr 5e-5, weight decay 0.01, linear warmup, max 12 epochs with early stopping (patience 3); best checkpoint at epoch 5 (validation loss 0.0248).

Results — SROIE test set (347 receipts)

Field	F1 (exact)	F1 (fuzzy)
Company	0.17	0.78
Date	0.61	0.70
Address	0.07	0.91
Total	0.82	0.87
Macro	0.42	0.81

Numbers reflect the full inference pipeline (OCR → model → post-processing, including a regex date fallback). Address/company exact-F1 are bounded by OCR noise in both the input and the SROIE ground truth; fuzzy-F1 is the more representative figure for those fields.

Usage

This model expects pre-computed OCR words and normalized boxes (apply_ocr=False) plus the page image. The full pipeline and an interactive demo are available here:

• Demo: https://huggingface.co/spaces/Tanishq71/Receipt-Entity-Extractor
• Code: https://github.com/Tanishqarya17/Receipt-Entity-Extractor

Limitations

Specialized to SROIE-style receipts. Zero-shot transfer to other schemas/languages degrades sharply (near-zero macro-F1 on Indonesian CORD; 0.38 fuzzy macro on English WildReceipt). Date fields transfer best; structured fields like address and company do not. Not validated for production use.