screenbench 0.0.1

ScreenSuite

ScreenSuite

A comprehensive benchmaring suite for evaluating Graphical User Interface (GUI) agents (i.e. agents that act on your screen, like our Computer Agent) across areas of ability : perception, single-step and multi-step agentic behaviour.

This does not aim to compare agent implementations, only the MLLMs that power them: thus we propose only simple agent implementations based on smolagents.

GUI Agent Benchmarks Overview

Grounding/Perception Benchmarks

Data Source	Evaluation Type	Platform	Link
ScreenSpot	BBox + click accuracy	Web	HuggingFace
ScreenSpot v2	BBox + click accuracy	Web	HuggingFace
ScreenSpot-Pro	BBox + click accuracy	Web	HuggingFace
Visual-WebBench	Multi-task (Caption, OCR, QA, Grounding, Action)	Web	HuggingFace
WebSRC	Web QA	Web	HuggingFace
ScreenQA-short	Mobile QA	Mobile	HuggingFace
ScreenQA-complex	Mobile QA	Mobile	HuggingFace
Showdown-Clicks	Click prediction	Web	HuggingFace

Single Step - Offline Agent Benchmarks

Data Source	Evaluation Type	Platform	Link
Multimodal-Mind2Web	Web navigation	Web	HuggingFace
AndroidControl	Mobile control	Mobile	GitHub

Multi-step - Online Agent Benchmarks

Data Source	Evaluation Type	Platform	Link
Mind2Web-Live	URL matching	Web	HuggingFace
GAIA	Exact match	Web	HuggingFace
BrowseComp	LLM judge	Web	Link
AndroidWorld	Task-specific	Mobile	GitHub
MobileMiniWob	Task-specific	Mobile	Included in AndroidWorld GitHub
OSWorld	Task-specific	Desktop	GitHub

Cloning the Repository

Make sure to clone the repository with submodules required:

git clone --recurse-submodules [e-mail]:huggingface/geekagents.git

or

git submodule update --init --recursive # if you already cloned the repository. To run also when you pull branches to update the submodules

Requirements

• Docker
• Python >= 3.11
• uv

For multistep agent benchmarks, we need to spawn containers environment. To do so, you need KVM virtualization enabled. To check if your hosting platform supports KVM, run

egrep -c '(vmx|svm)' /proc/cpuinfo

on Linux. If the return value is greater than zero, the processor should be able to support KVM. Note: macOS hosts generally do not support KVM.

Installation

# Using uv (faster)
uv sync --extra submodules --python 3.11

If you encounter issues with evdev python package, you can try installing the build-essential package:

sudo apt-get install build-essential

Development

# Install development dependencies
uv sync --all-extras

# Run tests
uv run pytest

# Code quality
uv run pre-commit run --all-files --show-diff-on-failure

Running the benchmarks

#!/usr/bin/env python
import os
import json
from datetime import datetime
from dotenv import load_dotenv
from smolagents.models import InferenceClientModel, OpenAIServerModel, LiteLLMModel
from screensuite import registry
from screensuite.basebenchmark import EvaluationConfig

load_dotenv()

# Setup results directory
RESULTS_DIR = os.path.join(os.path.dirname(__file__), "results")
os.makedirs(RESULTS_DIR, exist_ok=True)

def run_benchmarks():
    # Get benchmarks to run
    # benchmarks = registry.list_all()
    benchmarks = registry.get_by_tags(
        tags=[
            "screenqa_short",
            "screenqa_complex",
            "screenspot-v1-click-prompt",
            "screenspot-v1-bounding-box-prompt",
            "screenspot-v2-click-prompt",
            "screenspot-v2-bounding-box-prompt",
            "screenspot-pro-click-prompt",
            "screenspot-pro-bounding-box-prompt",
            "websrc_dev",
            "visualwebbench",
            "android_control",
            "showdown_clicks",
            "mmind2web",
            "android_world",
            "osworld",
            "gaia_web",
        ]
    )

    for bench in benchmarks:
        print(bench.name)

    # Configure your model (choose one)
    model = InferenceClientModel(
        model_id="Qwen/Qwen2.5-VL-32B-Instruct",
        provider="fireworks-ai",
        max_tokens=4096,
    )

    # Alternative models:
    # model = OpenAIServerModel(model_id="gpt-4o", max_tokens=4096)
    # model = LiteLLMModel(model_id="anthropic/claude-sonnet-4-20250514", max_tokens=4096)
    # see smolagents documentation for more models -> https://github.com/huggingface/smolagents/blob/main/examples/agent_from_any_llm.py

    # Run benchmarks
    run_name = f"test_{datetime.now().strftime('%Y-%m-%d')}"
    max_samples_to_test = 200
    parallel_workers = 4
    osworld_env_config = OSWorldEnvironmentConfig(provider_name="docker")

    for benchmark in benchmarks:
        print(f"Running: {benchmark.name}")

        # Configure based on benchmark type
        config = EvaluationConfig(
            parallel_workers=parallel_workers,
            run_name=run_name,
            max_samples_to_test=max_samples_to_test
        )

        try:
            results = benchmark.evaluate(
                model,
                evaluation_config=config,
                env_config=osworld_env_config if "osworld" in benchmark.tags else None,
            )
            print(f"Results: {results._metrics}")

            # Save results
            with open(f"{RESULTS_DIR}/results_{run_name}.jsonl", "a") as f:
                entry = {"benchmark_name": benchmark.name, "metrics": results._metrics}
                f.write(json.dumps(entry) + "\n")

        except Exception as e:
            print(f"Error in {benchmark.name}: {e}")
            continue

if __name__ == "__main__":
    run_benchmarks()

OSWorld Google Tasks

To run OSWorld Google tasks, you need to create a Google account and a Google Cloud project. See OSWorld documentation for more details.

License

This project is licensed under the terms of the Apache License 2.0.