---
license: apache-2.0
language:
- fa
pipeline_tag: image-to-text
library_name: paddleocr
tags:
- ocr
- persian
- handwriting
- paddleocr
- pp-ocrv6
- ctc
datasets:
- Reza2kn/persian-handwriting-pages-3.69m
- Reza2kn/persian-printed-ocr-3.5m
---

# Bina 0.2 Rizeh

Bina 0.2 Rizeh is a compact Persian printed and handwriting recognizer fine-tuned from the
`PP-OCRv6_medium_rec` architecture. It uses PaddleOCR's CTC recognition head and ships both the
exported inference model and the resumable training checkpoint from the Stage-8 release.

## Important model boundary

The neural checkpoint recognizes **text-line or text-region crops**. It is not, by itself, a
standalone full-page detector. The full-page benchmark below used the included adaptive75
**two-layer pipeline**: page-region detection and template isolation, CTC recognition of the
remaining regions, then reading-order assembly. The scripts used for that exact run are preserved
under `pipeline/` and the benchmark records are under `benchmark/`.

## Persian OCR Triple Threat results

All 13,338 frozen benchmark rows were scored, with zero missing predictions. Lower is better.
Triple Threat is `20% WER + 20% CER + 60% S3`; Overall is a 50/50 macro-average of the printed and
handwriting splits.

| Split | Pages | WER | CER | S3 | Essential error | Triple Threat |
|---|---:|---:|---:|---:|---:|---:|
| Handwriting | 6,669 | 19.0833 | 13.6399 | 13.7015 | 14.7567 | 14.7656 |
| Printed | 6,669 | 8.9257 | 2.3535 | 9.8512 | 12.0774 | 8.1666 |
| Equal-split overall | 13,338 | 14.0045 | 7.9967 | 11.7764 | 13.4171 | **11.4661** |

Leaderboard: [Persian OCR Triple Threat](https://huggingface.co/spaces/Reza2kn/PersianOCR-TripleThreat)

The release did **not** reach the single-digit handwriting gate. Printed Triple Threat is already
single digit; handwriting remains the principal limitation.

## Training snapshot

- Stage-8 training mix: 2,198,483 recognition crops.
- Mix composition: 1,000,000 clean handwriting, 1,000,000 printed, and 198,483 native/form crops.
- Best handwriting validation normalized edit distance: 0.9671727.
- Best printed validation normalized edit distance: 0.9071054.
- Optimizer state was preserved from Stage 4.
- The planned Stage-8 schedule completed without an overfitting stop signal.

The source pages came from
[`Reza2kn/persian-handwriting-pages-3.69m`](https://huggingface.co/datasets/Reza2kn/persian-handwriting-pages-3.69m)
and
[`Reza2kn/persian-printed-ocr-3.5m`](https://huggingface.co/datasets/Reza2kn/persian-printed-ocr-3.5m),
with line/region preparation and native form replay described by `data/MIX_MANIFEST.json`.

## Inference artifacts

The exported Paddle model is available both at repository root and under `inference/`:

- `inference.json`
- `inference.pdiparams`
- `inference.yml`

With PaddleOCR 3.x, a recognition crop can be run with the custom model directory:

```python
from paddleocr import TextRecognition

recognizer = TextRecognition(model_dir="./inference")
for result in recognizer.predict(input="persian-line.png", batch_size=1):
    result.print()
```

For page OCR, use a detector/layout stage first and pass its ordered text crops to this recognizer.
The bundled `pipeline/` scripts record the exact full-page benchmark workflow; their original
absolute cluster paths must be adapted to a new machine.

## Repository layout

- `inference/`: exported Paddle inference graph, parameters, and configuration.
- `checkpoint/`: resumable best checkpoint, optimizer state, configs, and eval-gate state.
- `pipeline/`: detection, recognition, assembly, benchmark, and launch scripts.
- `benchmark/`: exact split metrics, logs, and run summaries.
- `data/MIX_MANIFEST.json`: exact Stage-8 train/eval composition.
- `RELEASE.json`: immutable release identity and headline metrics.
- `SHA256SUMS`: checksums for the verified packaged release.

## Release identity

This public model name is **Bina 0.2 Rizeh**. The immutable packaging metadata retains the internal
release name `Bina-0.2-RizehPizeh` and release ID `stage8-gate42-20260726` so it remains traceable to
the exact trained artifact.