Create README.md
#1
by
cogito233
- opened
README.md
ADDED
|
@@ -0,0 +1,38 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
datasets:
|
| 4 |
+
- TIGER-Lab/BrowserAgent-Data
|
| 5 |
+
language:
|
| 6 |
+
- en
|
| 7 |
+
base_model:
|
| 8 |
+
- Qwen/Qwen2.5-7B-Instruct
|
| 9 |
+
metrics:
|
| 10 |
+
- success_rate
|
| 11 |
+
- trajectory_f1
|
| 12 |
+
tags:
|
| 13 |
+
- agent
|
| 14 |
+
- browser
|
| 15 |
+
- web
|
| 16 |
+
- sft
|
| 17 |
+
---
|
| 18 |
+
|
| 19 |
+
## Model
|
| 20 |
+
We release the **SFT (Supervised Fine-Tuned)** model used in **BrowserAgent**, based on `Qwen/Qwen2.5-7B-Instruct`.
|
| 21 |
+
This model learns structured web-browsing behaviors—such as **click, type, scroll, read, submit**—from human-style demonstrations and produces **schema-constrained action sequences** for browser environments.
|
| 22 |
+
|
| 23 |
+
- **Base:** Qwen2.5-7B-Instruct
|
| 24 |
+
- **Objective:** Next-token prediction on normalized, schema-validated browsing trajectories
|
| 25 |
+
- **Format:** JSON-like structured actions (compatible with BrowserAgent runtime)
|
| 26 |
+
|
| 27 |
+
## Data
|
| 28 |
+
The SFT data includes:
|
| 29 |
+
- Human and assisted browsing demonstrations
|
| 30 |
+
- Canonicalization under a unified action schema
|
| 31 |
+
- Filtering and de-duplication to ensure validity and safety
|
| 32 |
+
|
| 33 |
+
## Code
|
| 34 |
+
<https://github.com/TIGER-AI-Lab/BrowserAgent>
|
| 35 |
+
|
| 36 |
+
## Sample Usage
|
| 37 |
+
```bash
|
| 38 |
+
hf download TIGER-Lab/BrowserAgent-SFT --local-dir ./models/browseragent-sft --repo model
|