This evaluation benchmark is built from licensed Balochi printed books, including rare lithographed prints in Arabic-derived script. It provides a held-out evaluation set, kept separate from the training corpus, along with native-expert verified ground-truth transcriptions, to support evaluation of OCR and vision-language model (VLM) recognition performance.
Potential Use Cases
- Balochi OCR/VLM Recognition Accuracy Measurement: Model-generated recognition results on document images can be compared against the verified ground truth to measure character- and word-level recognition accuracy.
- Error Type Analysis on Rare Lithograph Prints: The dataset can be used to analyze the specific types of recognition errors that occur on challenging Balochi documents, including rare lithographed prints in Arabic-derived script.