Update README.md

README.md

@@ -1,3 +1,49 @@
----
-license: cc-by-4.0
----
+---
+license: cc-by-4.0
+---
+# RAG-Adapt Mini v0.1
+
+## Overview
+RAG-Adapt Mini is a small multi-hop retrieval dataset constructed from HotpotQA (distractor version).
+
+Each example contains:
+- Query
+- Reference Answer
+- Candidate document pool (BM25-generated hard negatives + gold docs)
+- Gold document indices
+
+## Intended Use
+- Retrieval evaluation (Recall@k, MRR)
+- Adaptive retrieval policies
+- Multi-hop RAG experimentation
+- Reinforcement learning for retrieval strategy selection
+
+## Format
+Each JSONL record contains:
+
+{
+  id: string,
+  query: string,
+  reference_answer: string,
+  doc_pool: [
+    { doc_id, text, source }
+  ],
+  gold_doc_indices: [int],
+  metadata: dict,
+  created_at: string,
+  provenance: dict
+}
+
+## Dataset Size
+- Train: 2000 examples
+- Validation: 200 examples
+- Documents per example: 15
+
+## Source
+Derived from:
+HotpotQA (Yang et al., 2018)
+
+## Limitations
+- Gold matching based on title-level alignment.
+- Negatives generated using BM25 only.
+- No dense retrieval used in this mini version.