AHCR-238: A Large-Scale Dataset for Amharic Handwriting Deep Learning

We introduce AHCR-238, a large-scale, publicly available dataset for Amharic handwritten character recognition comprising 130,900 character images across all 238 core classes of the Amharic syllabary (Fidel). Amharic, the official language of Ethiopia and the second most spoken Semitic language worldwide, uses the ancient Ethiopic script, whose high inter-class visual similarity poses unique challenges for recognition systems. Yet no standardized, publicly accessible benchmark exists for isolated Amharic characters. AHCR-238 addresses this gap by combining 59{,}500 handwritten samples from over 250 writers with 71{,}400 synthetically augmented samples generated via elastic distortion from 50+ Amharic font families. We formalize the preprocessing pipeline, define fixed train/validation/test splits (75/15/10\%), and establish baselines with two CNN architectures: a VGG-style series network (10.2M parameters) and a ResNet-inspired residual network (6.8M parameters). The residual network achieves a macro F1-score of 0.971 while reducing parameters by 33\%. Dataset, evaluation protocol, and baseline code are publicly released.