Link dataset to paper and GitHub repository

README.md

@@ -1,34 +1,37 @@
 ---
-pretty_name: CARE-Pro
 language:
 - zh
 - hi
 - es
 - en
+size_categories:
+- n<1K
 task_categories:
 - question-answering
 - text-generation
+pretty_name: CARE-Pro
 tags:
 - multilingual
 - cross-cultural
 - local-knowledge
 - cultural-knowledge
-size_categories:
-- n<1K
 ---
 
-# Introduction
+# CARE-Pro: Cross-lingual and Cross-cultural Evaluation Dataset
+
+[**Paper**](https://huggingface.co/papers/2605.25360) | [**GitHub**](https://github.com/Guochry/LRPO)
 
-- [GitHub repository](https://github.com/Guochry/LRPO)
+CARE-Pro (Cross-lingual and Cross-cultural evaluation for Realistic multilingual information needs) is an evaluation benchmark introduced in the paper "Learning to Route Languages for Multilingual Policy Optimization". 
 
-CARE-Pro is a **cross-lingual** and **cross-cultural** evaluation dataset for realistic multilingual information needs. 
-The dataset contains both *fine-grained insider regional knowledge*, which focus on region-specific facts and practices, and *cross-cultural questions*, which ask about another region from a foreign-language perspective.
-The current release contains 775 examples in Chinese, Hindi, and Spanish, with English translations provided when available.
+The dataset targets two settings that are often underrepresented in standard benchmarks:
+- **Fine-grained insider regional knowledge**: Questions focusing on local, city-, town-, or community-level facts and practices rather than broad country-level facts.
+- **Cross-cultural information seeking**: Questions where users ask about another region or culture from a foreign-language perspective.
 
+The current release contains 775 examples in Chinese (**zh**), Hindi (**hi**), and Spanish (**es**), with English (**en**) translations provided when available.
 
 ## Dataset Format
 
-Each item has the following fields, in order:
+Each item has the following fields:
 
 | Field | Description |
 | --- | --- |
@@ -40,10 +43,17 @@ Each item has the following fields, in order:
 | `region` | Region, country, city, or cultural context associated with the example. |
 | `type` | Either `local` or `cross-cultural`. |
 
+## Evaluation
 
+To evaluate on CARE-Pro, generate model responses for each question and compare them against the gold reference. The authors suggest using an LLM-as-a-judge approach (prompt available in the [GitHub repository](https://github.com/Guochry/LRPO)) to assign one of four labels: 
+- `CORRECT`
+- `CORRECT_BUT_WRONG_LANGUAGE`
+- `INCORRECT`
+- `NOT_ATTEMPTED`
 
-## Citation
+Only `CORRECT` is counted as correct for the final accuracy.
 
+## Citation
 
 ```bibtex
 @misc{lrpo,