--- license: other license_name: internal language: [ne, en] tags: [ocr, document-parsing, nepali] --- # ocr-tech-k2 (archive) This is a private archive of the hybrid Nepali and English OCR and document-structure toolkit (`ocr-tech` K2 branch), migrated on 16 Nov 2026 from `~/ocr/ocr-tech-k2` on the OCR server. It contains the source (`src/`, `scripts/`, `configs/`, `tests/`), 20 trained models in `runs/hf-*` (`model.safetensors`) together with the Paddle bundles, benchmark outputs and data. The project's own README (install, CLI, training recipes) is in [`README_PROJECT.md`](README_PROJECT.md). **Left out:** training state (`optimizer.pt`, `scheduler.pt`, `rng_state.pth`). The checkpoints load for inference and fine-tuning, but a run can't be resumed exactly where it stopped. **Format.** `outputs/` and `data/` hold many small files, so they are packed into ~1 GB tar shards, each with a sha256 index (`*.index.json`).