Add dataset card and paper link

README.md

@@ -1,3 +1,39 @@
----
-license: cc-by-4.0
----
+---
+license: cc-by-4.0
+task_categories:
+- text-generation
+---
+
+# eval_diverse_dataset
+
+This dataset is the shared evaluation corpus used in the paper [Measuring Maximum Activations in Open Large Language Models](https://huggingface.co/papers/2605.15572).
+
+## Dataset Summary
+
+The `eval_diverse_dataset` is a 5,000-sample multi-domain corpus designed to characterize the dynamic range of activations in modern open Large Language Models (LLMs). It serves as a unified benchmark to measure global and layerwise maxima across various model families (such as Qwen and Gemma), architectures, and training stages. Characterizing these activations is critical for researchers working on low-bit quantization, activation scaling, and stable LLM inference.
+
+## Resources
+
+- **Paper:** [Measuring Maximum Activations in Open Large Language Models](https://huggingface.co/papers/2605.15572)
+- **GitHub Repository:** [clx1415926/Max_act_llm](https://github.com/clx1415926/Max_act_llm)
+
+## Key Features
+
+- **Size:** 5,000 diverse samples.
+- **Purpose:** Measurement of maximum activation magnitude (`M = max |a|`) for deployment-oriented analysis.
+- **Coverage:** Multi-domain text to ensure robust characterization of outlier features and residual stream peaks across different model series.
+
+## Usage
+
+The dataset is intended to be used with the scripts provided in the official repository. It can be converted to the specific tokenizer encoding of various model families (e.g., Qwen, Gemma) to perform activation analysis using PyTorch hooks.
+
+## Citation
+
+```bibtex
+@article{chen2025measuring,
+  title={Measuring Maximum Activations in Open Large Language Models},
+  author={Chen, Luxuan and Tian, Han and Chen, Xinran and Kong, Rui and Wang, Fang and Jiamin Chen and Yuchen Li and Jiashu Zhao and Shuaiqiang Wang and Haoyi Xiong and Dawei Yin},
+  journal={arXiv preprint arXiv:2605.15572},
+  year={2025}
+}
+```