Add dataset card and metadata

README.md

@@ -0,0 +1,24 @@
+---
+task_categories:
+- question-answering
+license: apache-2.0
+tags:
+  - long-context
+  - reinforcement-learning
+  - question-answering
+  - docqa
+---
+
+# DocQA-RL-1.6K: A Reinforcement Learning Dataset for Long-Context Question Answering
+
+This dataset, `DocQA-RL-1.6K`, is a specialized reinforcement learning (RL) training dataset comprising 1.6K document question answering (DocQA) problems. These problems span three reasoning domains: mathematical, logical, and multi-hop reasoning, designed to challenge large language models (LLMs) in long-context scenarios. The dataset was used to train the QwenLong-L1 model, detailed in the following paper:
+
+[QwenLong-L1: Towards Long-Context Large Reasoning Models with Reinforcement Learning](https://huggingface.co/papers/2505.17667)
+
+**Dataset Composition:**
+
+*   **Mathematical Reasoning:** 600 problems from the DocMath dataset, requiring numerical reasoning across long and specialized documents.
+*   **Logical Reasoning:** 600 multi-choice questions synthesized using DeepSeek-R1, requiring logic analysis of real-world documents.
+*   **Multi-Hop Reasoning:** 400 examples sampled from MultiHopRAG and Musique, emphasizing cross-document reasoning.
+
+**Code:** The code used for creating and utilizing this dataset can be found on Github: [https://github.com/Tongyi-Zhiwen/QwenLong-L1](https://github.com/Tongyi-Zhiwen/QwenLong-L1)