Bengali CRNN Phase D (Mozhi 86K real books) CER=0.0110 WER=0.0299 (baseline 25.46% -> now 1.10%)

English vocab v2 74 classes incl apostrophe and colon

English CRNN v2 architecture

English CRNN v2 MJSynth+IIIT5K CER=0.0023 apostrophe_fix

Classifier v2 real MLT acc=96.19%

Classifier v2 architecture trained on real ICDAR MLT crops

English vocab 72 classes

English CRNN architecture

English CRNN CER=0.0001 epoch=10

Script classifier architecture conf_threshold=0.70

Script classifier acc=100.00% bn=100.00% en=100.00% threshold=0.7

Bengali Phase C checkpoint

Bengali Phase B checkpoint

Bengali Phase A checkpoint

Bengali vocab (148 classes)

Bengali CRNN architecture

Bengali CRNN Phase B CER=0.0014 WER=0.0038

Detector inference config v2

Detector v2 mAP50=0.9403 — rebuilt synthetic generator

BN recognizer train chunk 9

BN recognizer train chunk 8

BN recognizer train chunk 7

BN recognizer train chunk 6

BN recognizer train chunk 5

BN recognizer train chunk 4

BN recognizer train chunk 3

BN recognizer train chunk 2

BN recognizer train chunk 1

Delete files dataset/bn_recognizer/train/*.jpg with huggingface_hub

BN recognizer train chunk 1

Detection dataset YAML config

Synthetic dataset detection v1 — Bangladeshi fonts + augmentation

Synthetic dataset clf_en v1 — Bangladeshi fonts + augmentation

Synthetic dataset clf_bn v1 — Bangladeshi fonts + augmentation

Synthetic dataset en_recognizer v1 — Bangladeshi fonts + augmentation

Synthetic dataset bn_recognizer v1 — Bangladeshi fonts + augmentation

Update README.md

Update README.md

Synthetic Dataset used for Fine Tuning v1

Script classifier architecture

Script classifier v1 acc=99.82%

English model v2 CER=0.0000 ep=4 — fixed digits + bold caps

English fine-tuned v2 CER=0.0054 (function words + bold caps)

English vocab v2

English model architecture v2

English model retrain v2 CER=0.0001 (fixed preprocessing + digit training)

Update bengali/vocab.json

English vocab

English model architecture

English model CER=0.0000