Bina 0.2 Stage 6 portable recovery bundle

This bundle resumes the native-bbox corrective lane on an intact or replacement eight-GPU cluster.

Immutable inputs

Hugging Face artifact: Reza2kn/Bina-0.2-RizehPizeh/training-data/native-replay-v1.tar
Archive SHA-256: 2cb4b0507a1f6fc1fcf6619b679de5fd35a52e2ee8f55c57ce5ebf00c52ffc40
Accepted replay: 7,041 hash-matched pages and 206,917 samples
Fixed compositor: 6,984 pages and 205,289 samples
Archived compositor fallback: 57 pages and 1,628 samples
Excluded: 485 pages that matched neither known production compositor

Cluster prerequisites

The cluster root is /workspace/bina02-ppocr and must contain:

PaddleOCR/
venv/
bina02_stage5_hardforms.yml
watch_bina02_stage4_eval.py
run/train-stage4-clean2m-lr3e6/best_accuracy
data/corpus-v4/balanced

The existing RTL label-encoding patch must remain active in PaddleOCR.

Recovery sequence

Copy the three executable files in this directory to /workspace/bina02-ppocr.
Obtain a fresh final CDN URL for the Hugging Face replay archive.
Run fetch_native_replay_v2.sh "$SIGNED_CDN_URL" under supervisor.
Run run_bina02_stage6_v2.sh under supervisor.
Export the best accepted checkpoint.
Run the exact cached-Surya Persian OCR Triple Threat benchmark.
Publish only if printed S3 and handwriting S3 are both below 10 with zero missing predictions.

The training mix is approximately 30 percent exact native forms, 35 percent general handwriting replay, and 35 percent printed replay. It uses a 1e-6 learning rate, frozen batch normalization, trainable backbone blocks 4 and 5 plus CTC layers, aspect-ratio-preserving resize and padding, and no geometric RecAug.