metadata
license: cc-by-nc-sa-4.0
base_model: microsoft/layoutlmv3-base
datasets:
- dvgodoy/rvl_cdip_mini
metrics:
- accuracy
TrustDoc Document Classifier
LayoutLMv3-base fine-tuned for document type classification (16 RVL-CDIP classes), part of the TrustDoc project -- a document AI trust layer with calibrated confidence for human-in-the-loop review.
- Base model: microsoft/layoutlmv3-base (CC BY-NC-SA 4.0 -- non-commercial use only, inherited by this fine-tune)
- Training data: dvgodoy/rvl_cdip_mini, a 1% subset of RVL-CDIP (16 document classes)
- Validation accuracy: 85.8%
Limitations
RVL-CDIP has documented issues that affect this model: shortcut-feature bias (some predictions may rely on spurious per-page ID codes rather than content), ~8% label noise, train/test duplication, real PII in the source documents, and a tobacco-industry/1950s-2002-only domain -- generalization to modern documents is unproven.