Add dataset card with metadata and sample usage

README.md

@@ -0,0 +1,63 @@
+---
+task_categories:
+- image-text-to-text
+---
+
+# RationalRewards: Reasoning Rewards Scale Visual Generation Both Training and Test Time
+
+[Paper](https://huggingface.co/papers/2604.11626) | [Project Page](https://tiger-ai-lab.github.io/RationalRewards/) | [GitHub](https://github.com/TIGER-AI-Lab/RationalRewards)
+
+RationalRewards is a reasoning-based reward model and toolkit for visual generation. Instead of reducing human judgments to a single unexplained score, it generates explicit multi-dimensional critiques before scoring. This transforms reward models from passive evaluators into active optimization tools that support both train-time reinforcement learning and test-time prompt refinement.
+
+The datasets associated with this project include:
+- **SFT training data:** High-quality rationales used for Supervised Fine-Tuning.
+- **Preference evaluation data:** Benchmarking data for preference prediction (GenAIBench, MMRB2, ERBench).
+- **RL training data:** Data used for diffusion reinforcement learning.
+
+These datasets were constructed using **Preference-Anchored Rationalization (PARROT)**, a framework that recovers high-quality rationales from preference-only data through anchored generation, consistency filtering, and distillation.
+
+## Sample Usage
+
+The following code snippet demonstrates how to construct messages for a text-to-image (T2I) evaluation using RationalRewards templates:
+
+```python
+import base64
+
+def build_t2i_messages(prompt, image_bytes):
+    system_prompt = (
+        "You are an expert image generation evaluator. Your task is to evaluate "
+        "the quality of a generated image based on a user instruction. Afterwards, "
+        "you need to suggest how to refine the original user request to produce "
+        "better image generation (if any)."
+    )
+    image_b64 = base64.b64encode(image_bytes).decode()
+    user_instruction = f"User Instruction: {prompt}
+You are provided with one image:
+1. Generated Image <image>
+
+To do this, you must first assess the image on three critical aspects, provide justifications and absolute scores in 1-4 scale."
+    
+    parts = user_instruction.split("<image>")
+    content = [
+        {"type": "text", "text": parts[0]},
+        {"type": "image_url", "image_url": {"url": f"data:image/png;base64,{image_b64}"}},
+    ]
+    if len(parts) > 1:
+        content.append({"type": "text", "text": parts[1]})
+    
+    return [
+        {"role": "system", "content": system_prompt},
+        {"role": "user", "content": content},
+    ]
+```
+
+## Citation
+
+```bibtex
+@article{rationalrewards2026,
+  title   = {RationalRewards: Reasoning Rewards Scale Visual Generation Both Training and Test Time},
+  author  = {Haozhe Wang and Cong Wei and Weiming Ren and Jiaming Liu and Fangzhen Lin and Wenhu Chen},
+  journal = {arXiv:2604.11626},
+  year    = {2026}
+}
+```