About this project

BanglaNLP Hub is a community-maintained catalog of Bangla (Bengali) natural language processing resources, serving as a centralized directory for researchers and practitioners working with the Bangla language. The project aggregates papers, datasets, models, tools, and benchmarks across 13 NLP tasks, currently holding 712 papers, 63 datasets, 20 models, and 9 tools. The site is built as a static site using Astro and TypeScript, with no client-side framework — only small vanilla JavaScript islands for catalog search, theme toggle, mobile navigation, task tabs, table sorting, and BibTeX copying. All content is rendered from YAML files under the /data directory, with nothing hardcoded in the pages. A defining characteristic of the project is its strict data honesty policy. The catalog enforces several rules: no invented values (scores, citation counts, and BibTeX entries are never synthesized), leaderboards remain empty until sourced with proper citations, all published links are verified nightly, and entries are verified by hand against authoritative sources like the ACL Anthology, arXiv, and OpenAlex. The project has a history of removing fabricated entries inherited from an original design prototype, including a model that did not exist, a duplicate under an invented title, and a paper with no publication. The catalog supports shareable filters via URL parameters. Datasets can be filtered by query, task, license, and year; papers can be filtered by query, task, venue, and year. License filters group the catalog's license text into categories (open, nc, research, other) but do not verify or certify a dataset's license. The open category matches only MIT, Apache 2.0, CC0, ODC-BY, CC BY 4.0, and CC BY-SA 4.0. The project includes a validation system using Zod schemas that checks for missing or malformed fields, bad URLs, stale verification dates, duplicate IDs or links, and leaderboards referencing unknown datasets or papers. This validation runs in CI on every pull request and gates deployment. Contributions are welcome through two paths: opening a "Submit a resource" issue with a link (the fastest way in), or opening a PR directly by editing YAML files. The project credits all contributors in CONTRIBUTORS.md. Key areas needing help include reviewing paper task assignments (most papers were bulk-imported with task assignments by title heuristic), populating empty leaderboards (9 of 11 benchmarks ship empty), adding missing BibTeX entries, verifying dataset sizes, and adding new resources. The project also includes a design reference directory containing the original Claude Design prototype data from which the YAML was migrated, kept for provenance but not used at build time. The site code is licensed under MIT, while catalog entries link to third-party resources that carry their own licenses.