Skip to content

question about AgentBench evaluation #32

@EllenYiGe

Description

@EllenYiGe

I have a question regarding the AgentBench evaluation reported in Table 3 of the paper.

From the AgentBench results shown in the benchmark leaderboard (https://arxiv.org/abs/2308.03688, https://github.com/THUDM/AgentBench (Benchmarking Results)), models such as Qwen2.5-14B-Instruct achieve 17.6 on WebShop.

However, in Table 3 of your paper, the reported result for Qwen2.5-7B-Instruct on WebShop is 58.8, which appears significantly higher than the numbers shown in the AgentBench benchmark.

Could you clarify:
1. Whether the WebShop evaluation setting used in your Table 3 is the same as the original AgentBench benchmark?
2. If there are any modifications to the environment, evaluation protocol, or dataset?

Thank you very much for your time and help.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type
    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions