Datasetpublic

internlm/WildClawBench

WildClawBench Hard, practical, end-to-end evaluation for AI agents — in the wild. WildClawBench is an agent benchmark that tests what actually matters: can an AI agent do real work, end-to-end, without hand-holding? We drop agents into a live OpenClaw environment — the same open-source personal AI assistant that real users rely on daily — and throw 60 original tasks at them: clipping goal highlights from a football match, negotiating meeting times over multi-round… See the full description on the dataset page: https://huggingface.co/datasets/internlm/WildClawBench.

sourceHugging Facemitupdated 1mo agoView on Hugging Face

71likes9.1kdownloads

rooteval.yaml

eval.yaml16 linesDownload Raw Back to root

1name: WildClawBench2description: Hard, practical, end-to-end evaluation for AI agents — in the wild.3evaluation_framework: wildclawbench4tasks:5  - id: overall6    name: Overall7    config: default8    split: test9  - id: avg_time10    name: Avg Time (min)11    config: default12    split: test13  - id: avg_cost14    name: Avg Cost ($)15    config: default16    split: test