ToolSandbox

ToolSandbox is a benchmark for how well a model acts as a general-purpose agent. The official description says: "Agentic tool-use benchmark with multi-turn tool conversations". Below we describe in simple words what benchmarks of this type generally examine.

ToolSandbox measures how well a model acts as a general-purpose agent. In detail, it checks:

• solving long, multi-step tasks with many intermediate decisions
• using tools correctly (browsers, terminals, APIs, calendars, other apps)
• adapting when something unexpected happens mid-task
• staying on goal over long task horizons instead of wandering off

The score is accuracy: the percentage of tasks answered or solved correctly. 100% means every task was solved perfectly; higher is better.

Official methodology note: Agentic tool-use benchmark with multi-turn tool conversations. As evaluated on the LFM2.5-24B-A2B model card: vLLM, temperature = 0, max output tokens = 1024; user simulator GPT-4o-2024-08-06 with temperature = 0 and max output tokens = 1024.

What is the ToolSandbox benchmark?

ToolSandbox is a benchmark for how well a model acts as a general-purpose agent. The official description says: 'Agentic tool-use benchmark with multi-turn tool conversations'. Below we describe in simple words what benchmarks of this type generally examine.

What does ToolSandbox test?

ToolSandbox mainly measures how well a model acts as a general-purpose agent. Concretely, it checks: solving long, multi-step tasks with many intermediate decisions; using tools correctly (browsers, terminals, APIs, calendars, other apps); adapting when something unexpected happens mid-task.

How is the ToolSandbox score calculated?

The score is accuracy: the percentage of tasks answered or solved correctly. 100% means every task was solved perfectly; higher is better. The declared maximum score of this benchmark is 100.0.

Where do I get ToolSandbox data and scores?

The official source is https://huggingface.co/LiquidAI/LFM2.5-24B-A2B. The official benchmark page/reference is linked on this page. Every score on Open LLM Radar is also available as open data: download the scores JSON (http://open-llm-radar-api:8000/api/v1/benchmarks/toolsandbox/scores) or pull it directly from the Open LLM Radar API with any HTTP client.