exp40 — HTR-VT (ours) on exp35's data mix
Residual CNN + span masking + 6-layer transformer (dim 384) + linear CTC head, trained from scratch on HHD + SCE + Pinkas + 20% synthetic digits/punctuation lines. Input: grayscale line at 64 px height, mirrored (RTL to LTR). Charset: Mishkefet-v1. HHD validation CER (greedy): 4.70. Details in results.json.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support