Skip to content

Qwen3.6-Plus reproduces at 32.4 SR vs 45.7 reported #18

Description

@Levishery

Ran the canonical command on tag v0.1.0 (093a329), no code changes, data-v0.1.0:

python -m bench_env.run --split test --agent generic_v2 \ --env-url http://localhost:4173 \ --model-base-url https://dashscope.aliyuncs.com/compatible-mode/v1 \ --model-name qwen3.6-plus \ --parallel 8 --isolation pages --headless

L1 L2 L3 L4 SR FC
Reported 100.0 78.1 44.6 3.8 45.7 34.0
Ours 95.0 53.4 21.7 8.8 32.4 17.2

The profile differs qualitatively: we score higher on L4 but much lower on L2/L3, and FC is half yours (most of our failures are MAX_STEPS loops, not premature completion). Enabling enable_thinking only moved SR to 34.0, so that isn't the cause.

Question: Which dataset tag matches the leaderboard — the root README points to data-v0.1.0, while docs/getting-started.md points to data-v1.0.

Happy to share logs/trajectories.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions