metadata
license: cc-by-sa-4.0
NJ Open Data benchmark for evaluating RAG pipelines on real-world multi-modal and multi-format data pulled from the State of New Jersey's open data portal.
The NJ Open data corpus contains more than 85 mllion rows of structured CSVs, more than 25000 pages of rich PDFs, images, XMLs, json, and several other types of data pulled from over 700 sources and 23 state agencies.
The benchmark constitutes two datasets, Golden and Bronze as follows.
For both datasers, the questions were generated via asking an LLM judge to produce interesting questions that require pulling information from multiple sources to generate an answer.
| Benchmark | Metric | # Questions | Description |
|---|---|---|---|
| Golden | RAG Accuracy | 50 | Expert-verified answers that span multiple sources and modalities |
| Bronze | RAG Accuracy | 635 | LLM generated answers verified by an LLM-judge that span multiple sources and modalities |