Add README.md

README.md

@@ -0,0 +1,26 @@
+# Bina 0.2 - RizehPizeh
+
+Persian full-page OCR pipeline based on PP-OCRv6:
+
+1. Detect ordered Persian text regions.
+2. Recognize CTC-safe RTL word groups.
+3. Reconstruct lines and complete pages in reading order.
+
+## Current training state
+
+The initial PP-OCRv6 medium recognizer pilot reached a held-out normalized edit
+similarity of `0.1752708888` before its ephemeral Colab runtime reset. No model
+checkpoint survived that reset.
+
+The files under `training-kit/` reproduce data preparation, RTL segmentation,
+and evaluation-driven early stopping. Future best checkpoints are uploaded
+under `checkpoints/` so runtime resets do not erase training progress.
+
+## Data
+
+- `Reza2kn/persian-handwriting-pages-3.69m`
+- `Reza2kn/persian-printed-ocr-3.5m`
+- `Reza2kn/visualears-hardword-sentences` for rare-character coverage
+
+The Persian recognizer uses a character dictionary rather than a language-model
+tokenizer.