---
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.