agent-reward-bench
收藏资源简介:
AgentRewardBench是一个用于评估Web Agent轨迹自动评估指标的数据集,包含了多种agents与web相关的轨迹数据。数据集规模在1千到1万条轨迹之间,提供了annotations.csv文件用于数据标注。此数据集可用于研究和开发自动评估Web Agent表现的评价方法。
AgentRewardBench is a dataset designed for evaluating automated evaluation metrics for Web Agent trajectories. It includes trajectory data associated with various Web Agents, with the total number of trajectories ranging from 1,000 to 10,000. An annotations.csv file is provided for data annotation purposes. This dataset can be utilized for research and development of evaluation methods for automatically assessing Web Agent performance.
AgentRewardBench 数据集概述
基本信息
- 语言: 英语 (en)
- 标签: agents, web, llm
- 数据规模: 1K<n<10K
- 默认配置: annotations
- 数据文件: data/annotations.csv
数据集描述
AgentRewardBench 是一个用于评估Web Agent轨迹自动评估方法的数据集。相关研究论文为《AgentRewardBench: Evaluating Automatic Evaluations of Web Agent Trajectories》。
相关资源
- 代码: https://github.com/McGill-NLP/agent-reward-bench
- 论文: https://arxiv.org/abs/2504.08942
- 网站: https://agent-reward-bench.github.io
- 演示: https://huggingface.co/spaces/McGill-NLP/agent-reward-bench-demo
- 排行榜: https://huggingface.co/spaces/McGill-NLP/agent-reward-bench-leaderboard
数据集结构
数据集包含以下主要目录和文件结构:
trajectories/ ├── cleaned/ │ ├── assistantbench/ │ │ ├── GenericAgent-<LLM>/ │ │ │ ├── GenericAgent-<LLM>on<benchmark>.<split>/ │ │ │ | ├── <benchmark>.<split>.0.json │ │ │ | ├── ... │ │ │ ├── ... | | ├── ... │ ├── visualwebarena/ │ │ ├── ... │ ├── webarena/ │ │ ├── ... │ ├── workarena/ │ │ ├── ... ├── judgments/ │ ├── <benchmark>/ │ │ ├── GenericAgent-<LLM>/ │ │ │ ├── <judge>/ │ │ │ | ├── <benchmark>.<split>.0.json │ │ │ | ├── ... │ ├── ... ├── screenshots/ │ ├── <benchmark>/ │ │ ├── GenericAgent-<LLM>/ │ │ │ ├── <benchmark>.<split>.0/ │ │ │ | ├── screenshot_step_0.png │ │ │ | ├── ... │ │ │ ├── ... │ │ ├── ... │ ├── visualwebarena/ │ │ ├── ... │ ├── ...
加载方式
可以使用huggingface_hub库加载数据集:
python
from huggingface_hub import snapshot_download
snapshot_download( repo_id="McGill-NLP/agent-reward-bench", repo_type="dataset", local_dir="./trajectories/" )




