File size: 1,754 Bytes
b117fc2 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 | # 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
1. Copy the three executable files in this directory to
`/workspace/bina02-ppocr`.
2. Obtain a fresh final CDN URL for the Hugging Face replay archive.
3. Run `fetch_native_replay_v2.sh "$SIGNED_CDN_URL"` under supervisor.
4. Run `run_bina02_stage6_v2.sh` under supervisor.
5. Export the best accepted checkpoint.
6. Run the exact cached-Surya Persian OCR Triple Threat benchmark.
7. 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`.
|