bugsiesegal/form-field-labeling-florence

Formeze form-field labeler — LoRA adapter (lora-ocr-cleangt-v1e3, 2026-08)

Open-vocabulary form-field labeling adapter for the Formeze PDF pipeline, trained on the teacher-relabeled dataset revision (clean ground truth).

Recipe

• Base: microsoft/Florence-2-base @ 5ca5edf5bd017b9919c05d08aebef5e4c7ac3bac
• LoRA r=64, alpha=128, dropout 0.05, targets q/k/v/o_proj + fc1/fc2, PEFT 0.18.0
• Prompts: <REGION_TO_DESCRIPTION> + region tokens + k=12 nearest OCR tokens
• Data: bugsiesegal/form-fields-for-layout-labeled-pages @ b2d482e0… (Gemini-3.5-flash teacher relabel, verbatim-caption style, degenerate labels filtered; includes 9 real authentication-form templates), template-grouped train split, 2 epochs + 1 continuation epoch (fresh-optimizer, initialized from the epoch-2 adapter — checkpoint resume is blocked on torch 2.5.1 by transformers' CVE-2025-32434 guard)
• The immutable test split was never used for training or model selection.

Validation metrics (template-grouped validation split, 759 images, INT8 T4)

metric	value	policy floor
classification semantic accuracy	0.6048	≥ 0.60
end-to-end semantic accuracy	0.5202	≥ 0.45
identity / financial / medical label acc	0.658 / 0.661 / 0.653	≥ 0.60 each
authentication label acc (reported slice)	0.867 (24 GT)	non-gating
detector ECE (Platt-calibrated)	0.0100	≤ 0.12

Review-aid use only: labels are proposals requiring explicit user approval; no calibrated label confidence exists. See the Formeze model card and model-manifest.json for the confidence contract and release gating.