| --- |
| license: odc-by |
| task_categories: |
| - question-answering |
| language: |
| - en |
| pretty_name: memdoc |
| tags: |
| - rag |
| - conversational-memory |
| - agent-evaluation |
| --- |
| |
| # memdoc |
|
|
| A dataset for evaluating QA agents that retrieve from **both conversational memory and a document corpus**, while controlling for variance that comes from the question and its evidence themselves. |
|
|
| Each item is a question with 2-4 gold-evidence chunks partitioned across the two stores. The same questions and evidence can be presented as memory-only, document-only, or split across both. Because the question text and the underlying facts stay fixed, differences in agent behaviour can be attributed to **where evidence lives** and **how the agent uses the two stores**, rather than to a changing question mix. |
|
|
| The snapshot has **198 answerable questions**, three persona memory corpora, the MultiHop-RAG document collection, and an evidence-id → URL map. Questions and the document corpus are derived from [MultiHop-RAG](https://huggingface.co/datasets/yixuantt/MultiHopRAG) (ODC-BY). Memory sessions were generated for this work. |
|
|
| ## Files |
|
|
| | Path | Role | |
| | --- | --- | |
| | `experiment_questions.csv` | Answerable questions (JSON lists in evidence columns) | |
| | `memory_collection/persona_{1,2,3}.json` | Eval-persona memory corpora | |
| | `document_collection/multihop_corpus.jsonl` | Document store | |
| | `document_collection/evidence_id_to_url.jsonl` | Gold evidence → article URL | |
