Add dataset card and link to paper
#1
by nielsr HF Staff - opened
README.md
CHANGED
|
@@ -1,3 +1,62 @@
|
|
| 1 |
-
---
|
| 2 |
-
license: apache-2.0
|
| 3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
task_categories:
|
| 4 |
+
- image-text-to-text
|
| 5 |
+
---
|
| 6 |
+
|
| 7 |
+
# OS-SPEAR: A Toolkit for the Safety, Performance, Efficiency, and Robustness Analysis of OS Agents
|
| 8 |
+
|
| 9 |
+
This repository contains the dataset for the paper [OS-SPEAR: A Toolkit for the Safety, Performance, Efficiency, and Robustness Analysis of OS Agents](https://huggingface.co/papers/2604.24348).
|
| 10 |
+
|
| 11 |
+
**OS-SPEAR** is a comprehensive evaluation toolkit for **OS Agents**, designed to systematically assess their capabilities across four critical dimensions: Safety, Performance, Efficiency, and Robustness.
|
| 12 |
+
|
| 13 |
+
GitHub Repository: [Wuzheng02/OS-SPEAR](https://github.com/Wuzheng02/OS-SPEAR)
|
| 14 |
+
|
| 15 |
+
## 📊 Evaluation Benchmarks
|
| 16 |
+
|
| 17 |
+
OS-SPEAR consists of three core subsets:
|
| 18 |
+
|
| 19 |
+
| Dimension | Subset | Description |
|
| 20 |
+
| ----------- | ----------------------- | ----------- |
|
| 21 |
+
| Safety | `S-subset` | Diverse environment- and human-induced hazards. |
|
| 22 |
+
| Performance | `P-subset` | Curated via trajectory value estimation and stratified sampling. |
|
| 23 |
+
| Robustness | `R-subset` | Cross-modal disturbances applied to visual and textual inputs. |
|
| 24 |
+
| Efficiency | Derived from `P-subset` | Quantifies performance via temporal latency and token consumption. |
|
| 25 |
+
|
| 26 |
+
## 📦 Dataset Preparation
|
| 27 |
+
|
| 28 |
+
After downloading the dataset, ensure the folders are placed in your local toolkit directory as follows:
|
| 29 |
+
|
| 30 |
+
```
|
| 31 |
+
S-subset/
|
| 32 |
+
P-subset/
|
| 33 |
+
R-subset/
|
| 34 |
+
```
|
| 35 |
+
|
| 36 |
+
## 🚀 Usage
|
| 37 |
+
|
| 38 |
+
To evaluate an OS agent using the OS-SPEAR toolkit, you can configure the `eval/config.yaml` file:
|
| 39 |
+
|
| 40 |
+
```yaml
|
| 41 |
+
MODEL: <model_name> # choose from supported models
|
| 42 |
+
MODEL_PATH: <absolute_path_to_model>
|
| 43 |
+
DATA_PATH: <absolute_path_to_OS-SPEAR>
|
| 44 |
+
LOG_PATH: <path_to_save_logs>
|
| 45 |
+
|
| 46 |
+
TEST_S: true # Safety
|
| 47 |
+
TEST_P: true # Performance
|
| 48 |
+
TEST_R: true # Robustness
|
| 49 |
+
```
|
| 50 |
+
|
| 51 |
+
Then, run the evaluation script:
|
| 52 |
+
|
| 53 |
+
```bash
|
| 54 |
+
cd eval
|
| 55 |
+
python run.py
|
| 56 |
+
```
|
| 57 |
+
|
| 58 |
+
Finally, generate the evaluation report:
|
| 59 |
+
|
| 60 |
+
```bash
|
| 61 |
+
python report/evaluate.py
|
| 62 |
+
```
|