Is a Generic Dataset and Foundation VLM for Arabic HTR Worth It? Lessons from AMIDDA

Which handwritten text recognition (HTR) strategy works best for Arabic scripts, for which use case, and at what cost? We address this question with AMIDDA, an open dataset that compiles eight Arabic-script corpora into 66,779 real line images spanning Maghribi, Oriental, lithographed and modern scripts, with no synthetic data. On its test split we compare six strategies, from training-free in-context learning (ICL) with generalist vision-language models (VLMs) to specialised CRNN networks. Specialised CRNNs remain the most accurate option for production, while ICL reaches competitive character error rates (below 8 % on modern KHATT, 10.6 % on historical RASM) from minimal input, a few line examples or the plain transcription of one reference page. Backed by a cost evaluation, ICL stands as a credible bootstrapping mechanism: it produces first-pass annotations at scale that a human corrector can turn into the training data of a specialised model, and a Mixture-of-Experts post-correction step adds a low-cost lift when available models generalise poorly. We release the dataset, a Qwen3.5-4B LoRA foundation model covering the eight corpora, the ICL and MoE code, and a web application exposing these pipelines.