Reference companion MathNet-Retrieve dataset (quick start, tasks callout, links)

README.md

@@ -309,6 +309,12 @@ print(row["competition"], row["country"])
 print(row["problem_markdown"])
 for img in row["images"]:
     img.show()  # PIL image — renders inline in the HF viewer
+
+# Companion retrieval benchmark — MathNet-Retrieve (BEIR/MTEB format)
+# tier in {"easy", "medium", "hard"}
+corpus  = load_dataset("ShadenA/MathNet-Retrieve", "hard", split="corpus")   # {_id, text}
+queries = load_dataset("ShadenA/MathNet-Retrieve", "hard", split="queries")  # {_id, text}
+qrels   = load_dataset("ShadenA/MathNet-Retrieve", "hard", split="qrels")    # {query-id, corpus-id, score}
 ```
 
 ## Overview
@@ -328,6 +334,10 @@ MathNet spans **47 countries**, **17 languages**, and **two decades** of competi
 
 Even state-of-the-art reasoners remain challenged: **78.4% (Gemini-3.1-Pro)** and **69.3% (GPT-5)** on `MathNet-Solve-Test`. Embedding models struggle with equivalence retrieval (Recall@1 under 5% for all tested models), and RAG gains are highly sensitive to retrieval quality — expert retrieval lifts DeepSeek-V3.2-Speciale to **97.3%** on `MathNet-RAG`.
 
+> **Companion dataset:** Task II (Math-Aware Retrieval) is released separately as
+> [`ShadenA/MathNet-Retrieve`](https://huggingface.co/datasets/ShadenA/MathNet-Retrieve) —
+> equivalent-problem retrieval graded across `easy` / `medium` / `hard` difficulty tiers.
+
 
 ## How MathNet compares to existing math benchmarks
 
@@ -455,6 +465,7 @@ If you are a rightsholder with a concern, please open an issue or email [shaden@
 
 - 🌐 **Website & paper:** <https://mathnet.mit.edu>
 - 🔭 **Browse all 30K problems:** <https://mathnet.mit.edu/explorer.html>
+- 🧩 **Retrieval benchmark:** <https://huggingface.co/datasets/ShadenA/MathNet-Retrieve>
 - ✉️ **Contact:** [shaden@mit.edu](mailto:shaden@mit.edu)
 
 <p align="center"><sub>© 2026 Massachusetts Institute of Technology · MathNet · ICLR 2026</sub></p>