agentboard / README.md
YANG-Cheng's picture
Update README.md
c662f95 verified
metadata
license: gpl-2.0
configs:
  - config_name: alfworld
    data_files:
      - split: test
        path:
          - data/alfworld/test.jsonl
  - config_name: scienceworld
    data_files:
      - split: test
        path:
          - data/scienceworld/test.jsonl
  - config_name: babyai
    data_files:
      - split: test
        path:
          - data/babyai/test.jsonl
  - config_name: jericho
    data_files:
      - split: test
        path:
          - data/jericho/test.jsonl
  - config_name: pddl
    data_files:
      - split: test
        path:
          - data/pddl/test.jsonl
  - config_name: webarena
    data_files:
      - split: test
        path:
          - data/webarena/test.jsonl
  - config_name: webshop
    data_files:
      - split: test
        path:
          - data/webshop/test.jsonl
  - config_name: tool-query
    data_files:
      - split: test
        path:
          - data/tool-query/test.jsonl
  - config_name: tool-operation
    data_files:
      - split: test
        path:
          - data/tool-operation/test.jsonl
language:
  - en
tags:
  - Embodied AI
  - Game
  - Web
  - Tool
size_categories:
  - 1K<n<10K
task_categories:
  - text-generation
pretty_name: AgentBoard

AgentBoard: An Analytical Evaluation Board of Multi-turn LLM Agents

This is the official dataset repository of AgentBoard.

1. Data Overview

AgentBoard is composed of 9 diverse tasks which can be divided into 4 types, including Embodied AI, Game, Web, and Tool:

Embodied AI Game Web Tool
  • AlfWorld
  • ScienceWorld
  • BabyAI
  • Jericho
  • PDDL
  • WebShop
  • WebArena
  • Tool-Query
  • Tool-Operation

And statistics of the evaluation data of 9 environments are as follows:

AlfWorld ScienceWorld BabyAI Jericho PDDL WebShop WebArena Tool-Query Tool-Operation
#Environment 134 90 112 20 60 251 245 60 40
#Turn 6 15 10 20 20 3 25 5 6
#Action Space 13 21 8 150 8 2 12 15 16
#Context Length 900 2800 1800 1500 2700 1200 15000 2100 4300
Progress Rate subgoal subgoal subgoal subgoal match match match subgoal subgoal/match
#Avg. Subgoals 3 5 4 6 6 4 6 5 5
Hard/Easy Cutoff 3 3 3 4 6 1 4 4 4

To help researchers quickly understand evaluation data of each task, we provide Dataset Viewer at Huggingface Dataset: πŸ€— AgentBoard.

Note: Please download the dataset from the link provided below for the reason that the data in Dataset Viewer is not complete.

2. Download Link

You can download the whole evaluation data by running the following command:

wget https://huggingface.co/datasets/hkust-nlp/agentboard/resolve/main/data.tar.gz

Please uncommpress the file and move the data to AgentBoard/data.

cd AgentBoard
mkdir data
tar -zxvf data.tar.gz

The file structure of evaluation data is as follows:

Click to expand the file structure
data
β”œβ”€β”€ alfworld
β”‚   β”œβ”€β”€ alfred.pddl # additional data for alfworld
β”‚   β”œβ”€β”€ alfred.twl2 # additional data for alfworld
β”‚   β”œβ”€β”€ json_2.1.1  # additional data for alfworld
β”‚   └── test.jsonl
β”œβ”€β”€ babyai
β”‚   └── test.jsonl
β”œβ”€β”€ jericho
β”‚   β”œβ”€β”€ test.jsonl
β”‚   └── z-machine-games-master  # additional data for jericho
β”œβ”€β”€ pddl
β”‚   └── test.jsonl
β”œβ”€β”€ scienceworld
β”‚   └── test.jsonl
β”œβ”€β”€ tool-operation
β”‚   └── test.jsonl
β”œβ”€β”€ tool-query
β”‚   β”œβ”€β”€ academia  # additional data for academia tool
β”‚   └── test.jsonl
β”œβ”€β”€ webarena
β”‚   └── test.jsonl
└── webshop
    └── test.jsonl

3. Data Fields

We take an instance from the ScienceWorld task as an example to illustrate the data fields of evaluation data.

{
  "task": "scienceworld",
  "id": 0,
  "goal": "Your task is to find the animal with the longest life span.  The animals are in the 'outside' location.  Focus on the animal with the longest life span.",
  "subgoals": ["You move to the outside.", "You focus on the crocodile egg."],
  "difficulty": "easy",
  "additional_info": {"var": 5, "env_name": "lifespan-longest-lived"}
}

Details of the data fields are as follows:

Field Name Description
task The task name of the example, e.g. alfworld, babyai, jericho, pddl, scienceworld, tool-operation, tool-query, webarena, webshop.
id The id of the example.
goal The goal of the example.
subgoals The subgoals of the example.
difficulty The difficulty of the example, e.g. easy, hard.
additional_info The additional information of the example, each example has its own additional information.

4. Citation

@misc{ma2024agentboard,
      title={AgentBoard: An Analytical Evaluation Board of Multi-turn LLM Agents}, 
      author={Chang Ma and Junlei Zhang and Zhihao Zhu and Cheng Yang and Yujiu Yang and Yaohui Jin and Zhenzhong Lan and Lingpeng Kong and Junxian He},
      year={2024},
      eprint={2401.13178},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}