Towards More Advanced QA Datasets 🔎 A big issue with evaluating RAG is that the datasets are oftentimes too simple - the questions are derived from a single chunk, and the source corpus is a single document. @tjaffri just contributed a fantastic dataset that addresses some of these concerns: a Multi-Doc SEC 10Q Dataset. The questions range in difficulty: ✅ Single-Doc, Single-Chunk RAG: Questions where answer is in a contiguous region of a single-doc. ✅ Single-Doc, Multi-Chunk RAG: Questions where answer can be found by combining multiple non-contiguous regions of a single doc. ✅ Multi-Doc RAG: Questions where answer can be found through combining multiple non-contiguous regions of multiple docs. Check it out! llamahub.ai/l/llama_datasets…