README.md

---
license: apache-2.0
language:
- zh
- en
tags:
- browser-agent
- web-automation
- benchmark
- evaluation
size_categories:
- n<1K
task_categories:
- question-answering
- text-generation
configs:
- config_name: LexBench-Browser
  data_files:
  - split: All
    path: LexBench-Browser/task.jsonl
---

# LexBench-Browser

LexBench-Browser is a benchmark for evaluating AI browser agents on real-world web tasks across 50+ Chinese and English websites — e-commerce, social, video, tools/education, finance, and more.

## Dataset Description

This release contains **210 tasks** spanning common multi-step browsing flows. None of the tasks gate on user login, so they can be benchmarked end-to-end without manual sign-in.

### Dataset Statistics (v3.1, 2026-04-29)

| Attribute | Value |
|-----------|-------|
| Total tasks | 210 |
| Version | 3.1 |
| Languages | Chinese (137), English (73) |
| `risk_control=true` tasks | 21 |

#### Task type

| Type | Count | Description |
|------|-------|-------------|
| T1   | 166   | Information retrieval |
| T2   | 44    | Website operations |

#### Reasoning type

| Reasoning | Count |
|-----------|-------|
| single_step  | 117 |
| multi_step   | 70  |
| deep_analysis | 23 |

#### Domain Distribution

| Domain | Count |
|--------|-------|
| finance_gaming   | 44 |
| video_platform   | 42 |
| tools_education  | 40 |
| general          | 34 |
| social_lifestyle | 26 |
| ecommerce        | 23 |
| gaming           | 1  |

## Download

### Using Hugging Face Datasets

```python
from datasets import load_dataset

dataset = load_dataset("Lexmount/LexBench-Browser")
all_tasks = dataset["All"]
print(all_tasks[0])
```

### For Users in China (Mirror)

```python
import os
os.environ['HF_ENDPOINT'] = 'https://hf-mirror.com'

from datasets import load_dataset
dataset = load_dataset("Lexmount/LexBench-Browser")
```

Or set the variable before running:

```bash
export HF_ENDPOINT=https://hf-mirror.com
python your_script.py
```

### Manual Download

```bash
pip install huggingface_hub
huggingface-cli download Lexmount/LexBench-Browser --repo-type dataset --local-dir ./data
```

For China users:

```bash
export HF_ENDPOINT=https://hf-mirror.com
huggingface-cli download Lexmount/LexBench-Browser --repo-type dataset --local-dir ./data
```

## Task Format

```json
{
  "id": 1,
  "query": "Task description",
  "task_type": "T1",
  "reasoning_type": "multi_step",
  "domain": "ecommerce",
  "difficulty": "medium",
  "login_required": false,
  "login_type": "",
  "risk_control": false,
  "risk_control_types": [],
  "target_website": "www.example.com",
  "language": "zh",
  "website_region": "zh",
  "reference_answer": {
    "steps": ["Step 1", "Step 2"],
    "key_points": ["Key point 1"],
    "common_mistakes": ["Common mistake 1"],
    "scoring": {
      "total": 100,
      "items": [{"name": "...", "score": 30, "description": "..."}]
    }
  },
  "score_threshold": 60,
  "robustness_tags": ["chinese_rendering", "data_extraction"]
}
```

### Field descriptions

- **task_type**: `T1` (information retrieval) or `T2` (website operations).
- **reasoning_type**: `single_step` | `multi_step` | `cross_platform` | `deep_analysis`.
- **domain**: business domain (`ecommerce`, `social_lifestyle`, `video_platform`, `tools_education`, `finance_gaming`, `general`, `gaming`).
- **difficulty**: `easy` | `medium` | `hard`.
- **language**: task description language (`zh` for Chinese, `en` for English).
- **website_region**: target website region (`zh` for Chinese sites, `en` for international sites).
- **login_required**: whether the task is login-gated (always `false` in this release).
- **risk_control** / **risk_control_types**: whether the target site applies anti-crawl / risk-control mechanisms (CAPTCHA, slider verification, anti-bot, rate-limiting).
- **target_website**: a single hostname, or for multi-site tasks, hostnames separated by `+` (e.g. `movie.douban.com + www.imdb.com`).
- **robustness_tags**: ordered, deduplicated list across 6 categories / 16 tags (popup interference, sequence complexity, content dynamics, anti-crawl, localization, complex interaction).

> **Migration note (v2.x → v3.x)**: the per-record `scenario_tier` field has been removed and the per-tier files (`l1.jsonl`, `l2.jsonl`, `l3-api.jsonl`, `l3-security.jsonl`, `tasks.jsonl`) have been replaced by a single `task.jsonl`. Slice the data with `login_required`, `domain`, or `risk_control` instead.

## License

Apache 2.0

## Citation

```bibtex
@misc{lexbench-browser-2026,
  title={LexBench-Browser: A Benchmark for Web Browsing AI Agents},
  author={Lexmount},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/Lexmount/LexBench-Browser}
}
```