byt5-sanskrit-ocr

ByT5-Sanskrit-OCR — Post-OCR Correction for Sanskrit/Devanagari

A byte-level ByT5-small fine-tuned to correct OCR'd Sanskrit/Devanagari text — dropped vowel signs, confused nasals (anusvara vs chandrabindu), split conjuncts, visually-similar consonants (व/ब, श/ष/स), and danda errors (॥ becoming a plain |). Trained on synthetic OCR noise from a linguistically-grounded Devanagari corruption engine. Built on a free Colab T4.

• Input: noisy OCR Sanskrit, prefixed with correct:
• Output: corrected Devanagari Unicode
• Base: google/byt5-small (Apache 2.0)
• Code & report: https://github.com/tusharislampure29/sanskrit-ocr-correction

Why byte-level (ByT5)?

mT5's subword tokenizer fragments 89.9% of Sanskrit words and fragments more as the input degrades. ByT5 is byte-level: no Devanagari input can produce an <unk> and the decoder can emit any Unicode sequence — exactly what a corrector of broken glyphs needs. The cost is ~3× longer sequences, acceptable for short verses. (Full analysis in the repo.)

Usage

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
tok = AutoTokenizer.from_pretrained("tusharislampure29/byt5-sanskrit-ocr")
model = AutoModelForSeq2SeqLM.from_pretrained("tusharislampure29/byt5-sanskrit-ocr")

noisy = "कर्मण्येवाधिकारस्त मा फलषु कदाचन |"   # dropped matras + danda-as-pipe
ids = tok("correct: " + noisy, return_tensors="pt").input_ids
print(tok.decode(model.generate(ids, max_length=256, num_beams=4)[0], skip_special_tokens=True))

Results

Held-out test set (1,800 lines; split by clean line, zero overlap with training), before vs after:

A strong word-level corrector: it more than halves the word error rate and returns 25% of lines exactly correct. CER improves most on heavily-degraded input (0.122 → 0.074, −40%); on already-light noise it can over-correct at the character level (CER 0.047 → 0.057) while still cutting WER (0.363 → 0.192). Plus a per-error-family taxonomy eval (which error types it fixes best) — see the repo.

Training

ByT5-small, seq2seq, prefix correct:, max len 384, lr 3e-4, effective batch 16 (4 × grad-accum 4), 3 epochs, bf16 on T4 (fp16 is NaN-unstable for T5/ByT5), best-by-validation-CER checkpoint (val CER 0.033). Dataset: tusharislampure29/sanskrit-ocr-correction.

Limitations

Trained on synthetic OCR noise — strongest on the modelled error families; novel scanner artifacts are out-of-distribution. Coverage is classical Sanskrit (Ayurveda/Yoga/Gita/subhashitas). For production manuscript pipelines, pair with a human-review/abstain step.

License & citation

Apache 2.0. Built for the ImmverseAI AI/ML assignment and released openly. Author: Tushar Islampure (github.com/tusharislampure29).