EyM Sanskrit OCR

Sanskrit-Qwen2.5-VL-7B / TrOCR-Devanagari ✅ Live — "Parinamika" server slot

Handwriting-trained neural models, served via ocr_server.py (loads from Hugging Face on first run, no manual download). Set EYM_NEURAL_ENGINE=qwen for the Sanskrit-fine-tuned Qwen2.5-VL (needs a GPU), or =trocr for the lighter TrOCR-Devanagari (CPU-friendly). This is the strongest option here specifically for cursive manuscript hands — the other engines below are printed-text or classical-LSTM models.

Shreeshrii/tesstrain-Sanskrit-IAST ✅ Live — "Sanskrit HTR" engine

Real, currently-hosted Devanagari Tesseract LSTM models, loaded directly by Tesseract.js from GitHub raw at OCR time — no server needed. Two variants selectable via the model dropdown:

Note: the originally-suggested tessdata4alpha repo (sa4.traineddata) and its CER 1.86%/WER 1.44% result no longer exist on GitHub — this repo is the maintainer's current, actively-hosted successor for Sanskrit Devanagari. Still classical LSTM Tesseract, not a handwriting specialist — use the neural engine above for cursive manuscripts when possible.

ihdia/sanskrit-ocr ⏸ Needs a server

IndicOCR-v2 / CNN-RNN, published CER/WER of 3.86%/13.86% and 3.77%/14.38% — but benchmarked on printed classical Sanskrit documents, not handwriting. The repo ships a real pretrained TensorFlow SavedModel, but its original serving path needs a separate TensorFlow-Serving binary plus tesserocr; wiring it in as a simple Flask /ocr endpoint (loading the SavedModel directly) is real but non-trivial extra work. Ask if you want this drafted as a fifth engine for printed sections.

Shreeshrii/imagessan 📄 Data only

Ground-truth Sanskrit images and transcriptions for evaluation — not a model, so there's nothing to run OCR "from" here. Its real use is as a benchmark set: run each engine above against these images to get real CER/WER numbers for EyM's manuscripts.