danieleschmidt/vislang-ultralow-resource

Dataset builder and training framework for visual-language models in ultra-low-resource languages. Scrapes humanitarian reports, OCRs infographics, and creates aligned vision-language instruction datasets following the Masakhane visual-instruction corpus approach (2025). # VisLang Ultra-Low Resource Dataset builder and training framework for visual-language models in ultra-low-resource languages. ## Features - HumanitarianReportScraper: Fetch UN OCHA humanitarian reports via ReliefWeb API - OCRPipeline: pytesseract-based text extraction with graceful fallback - AlignmentBuilder: Pair image regions with caption sentences via word overlap - DatasetExporter: Export datasets in HuggingFace-compatible JSONL format - DatasetStats: Vocabulary, image count, and language distribution statistics ## Install ``` pip install numpy requests # Optional: pip install pytesseract Pillow langdetect ``` ## Usage ```python from vislang.scraper import HumanitarianReportScraper from vislang.stats import DatasetStats scraper = HumanitarianReportScraper() reports = scraper.fetch_reports(query="flood response", limit=5) stats = DatasetStats() result = stats.compute([{"caption": r["body"], "image_path": ""} for r in reports]) print(stats.report(result)) ``` ## Run Tests ``` ~/anaconda3/bin/python3 -m pytest tests/ -v ```