Arabic manuscript HTR

This is our fine-tuned version of HAFITH, not a model trained from scratch. It uses SigLIP2 NaFlex, a projection connector, and a TrOCR decoder with an Arabic tokenizer (698 million distinct parameters).

We adapted the released checkpoint to 81,672 manuscript line images and continued full-model training with R-Drop. The selected checkpoint is update 2,500. Exported recognition weights are lossless FP32 Safetensors shards; 8-bit storage was used only for optimizer states during training.

Local evaluation

Strict corpus scores on 10,173 held-out line crops, with matched BF16 inference, 512 patches, and a 128-token generation limit:

These are our local results, not the original published benchmark or the evaluation engineer's setup. On 32 Quran pages assembled from 481 cached crops, stripped CER was 17.02%. That test did not use the included layout pipeline. See evaluation.json for details. Training exposure, near duplicates, and some crop references remain incompletely audited.

Use

Download this repository, install requirements.txt, then:

from PIL import Image
from hafith_htr import Recognizer

reader = Recognizer(".", device="cuda", precision="fp32")
text = reader.recognize_line(Image.open("line.png"))

Use device="cpu" for CPU inference. T4 deployment defaults to FP32; FP16 requires validation. Long lines may reach the generation limit. No LLM correction is applied.

The layout/ assets come from HAFITH's page models; they were not trained by us. Recognition code retains the upstream MIT notice; weight provenance is in release.json and NOTICE.

This release replaces the earlier CNN–Transformer–CTC model. That version remains available at tag legacy-v2-cnn-ctc-20261027.