Add paper link, GitHub link, and metadata

README.md

@@ -1,27 +1,39 @@
 ---
 license: mit
+language:
+- en
+task_categories:
+- text-generation
 configs:
 - config_name: train
   data_files:
-    - split: train
-      path: train.jsonl
+  - split: train
+    path: train.jsonl
 - config_name: single_test
   data_files:
-    - split: train
-      path: single_test.jsonl
+  - split: train
+    path: single_test.jsonl
 - config_name: multiple_test
   data_files:
-    - split: train
-      path: multiple_test.jsonl
+  - split: train
+    path: multiple_test.jsonl
 ---
 
-## dataset structure
+# SE-Bench: Benchmarking Self-Evolution with Knowledge Internalization
+
+[**Paper**](https://huggingface.co/papers/2602.04811) | [**GitHub**](https://github.com/thunlp/SE-Bench)
+
+SE-Bench is a diagnostic environment designed to rigorously measure an agent's ability to **internalize** novel knowledge, which is a foundational capability for true self-evolution. 
+
+The benchmark obfuscates the NumPy library and its API documentation into a pseudo-novel package with randomized identifiers. Agents are trained to internalize this package and are then evaluated on simple coding tasks without access to documentation. This yields a clean setting where tasks are trivial with the new API doc but impossible for base models without it.
+
+## Dataset Structure
 
 The SE-Bench contains the following subsets (configurations). Please note that all data is loaded under the `train` split key regardless of the subset.
-- `train`:
-  - `train` train set
-  - `single_test`: single function test set
-  - `multiple_test`: multiple function test set
+
+- `train`: Training questions and materials (including API documentation) used to teach the model the pseudo-novel API.
+- `single_test`: Test set for single-function coding problems.
+- `multiple_test`: Test set for multiple-function composition coding problems.
 
 ## Usage
 
@@ -38,6 +50,8 @@ print(dataset)
 
 ### Load Single Function Test Set
 ```python
+from datasets import load_dataset
+
 dataset = load_dataset("jintailin/SE-Bench", "single_test")
 # Data is in dataset['train']
 print(dataset)
@@ -45,7 +59,27 @@ print(dataset)
 
 ### Load Multiple Function Test Set
 ```python
+from datasets import load_dataset
+
 dataset = load_dataset("jintailin/SE-Bench", "multiple_test")
 # Data is in dataset['train']
 print(dataset)
+```
+
+## Protocol
+To use SE-Bench correctly for measuring knowledge internalization:
+1. Train your model or agent using **only** the information provided in the `train` configuration.
+2. Evaluate on problems in the `test` configurations **without** providing the model access to documentation at inference time. 
+
+## Citation
+
+If you find SE-Bench useful in your research, please cite:
+
+```bibtex
+@article{yuan2025sebench,
+  title={SE-Bench: Benchmarking Self-Evolution with Knowledge Internalization},
+  author={Jiarui Yuan and Tailin Jin and Weize Chen and Zeyuan Liu and Zhiyuan Liu and Maosong Sun},
+  journal={arXiv preprint arXiv:2602.04811},
+  year={2025}
+}
 ```