Add paper link, GitHub link, and task category to dataset card

README.md

@@ -1,3 +1,35 @@
----
-license: apache-2.0
----
+---
+license: apache-2.0
+task_categories:
+- text-generation
+tags:
+- reinforcement-learning
+- instruction-following
+- alignment
+---
+
+# HIR-16K
+
+This repository contains the dataset for the paper [Replay Failures as Successes: Sample-Efficient Reinforcement Learning for Instruction Following](https://huggingface.co/papers/2512.23457).
+
+[**GitHub**](https://github.com/sastpg/HIR) | [**Paper**](https://huggingface.co/papers/2512.23457)
+
+## Overview
+
+Reinforcement Learning (RL) for aligning Large Language Models (LLMs) often relies on sampling successful, high-quality responses. However, initial models often struggle with complex constraints, leading to sparse rewards.
+
+**Hindsight instruction Replay (HiR)** is a sample-efficient RL framework designed to address this. It employs a *select*-then-*rewrite* strategy to replay failed attempts as successes based on the constraints that were satisfied in hindsight. This approach frames the objective as dual-preference learning at both the instruction- and response-level, enabling efficient optimization using only binary reward signals.
+
+## Acknowledgement
+We thank [verl](https://github.com/volcengine/verl) for providing the open-source RL infrastructure. We also thank the developers of [Qwen](https://github.com/QwenLM) and [Llama](https://github.com/meta-llama) for their open-source models.
+
+## Citation
+
+```bibtex
+@article{hir2025replay,
+  title={Replay Failures as Successes: Sample-Efficient Reinforcement Learning for Instruction Following},
+  author={},
+  journal={arXiv preprint arXiv:2512.23457},
+  year={2025}
+}
+```