metadata
base_model: microsoft/Florence-2-base
library_name: peft
Formeze form-field labeler — LoRA adapter (lora-ocr-v2, 2026-08)
Open-vocabulary form-field labeling adapter for the Formeze PDF pipeline (YOLO detects field boxes; this adapter labels each region with free text).
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>+ Florence region tokens + k=12 nearest OCR tokens with positions (train_florence.py --use-ocr-context) - Data:
bugsiesegal/form-fields-for-layout-labeled-pages@62f3fded…, template-grouped train split, 227k field samples x 2 epochs, lr 1e-5 bf16, degenerate labels (UNKNOWN/N-A) filtered from targets - The immutable test split was never used for training or model selection.
Validation metrics (template-grouped validation split, 759 images, INT8 T4)
| metric | value |
|---|---|
| classification semantic accuracy (cosine >= 0.7) | 0.5835 |
| end-to-end semantic accuracy | 0.5019 |
| detection F1 (unchanged detector) | 0.8573 |
| high-risk identity / medical / authentication label acc | 0.768 / 0.647 / 0.765 |
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 full confidence contract and release gating.