Terminal-Bench Verified Dataset for Code Agent Evaluation

Terminal-Bench 2.0 Verified is a corrected version of a benchmark for evaluating AI code agents, addressing identified environment and instruction issues. The dataset was reviewed and modified by the organization zai-org, with the verified version released in February 2026. It includes updated Dockerfiles and instructions specifically to support the runtime of the Claude Code Agent.

Use Cases

• Benchmarking code agent performance like GLM-5 using the verified instruction set.
• Testing agent-environment interaction stability with the provided updated Dockerfiles.
• Comparing model outputs, such as from Step 3.5-Flash, against a standardized, corrected task suite.

Strengths

• Dataset underwent a comprehensive review to identify and fix various issues from the original Terminal-Bench 2.0.
• Includes specific fixes for environment setup and instruction clarity to support a named agent runtime.

Limitations

• The exact size, row count, and specific data columns are not provided in the description.
• The verification scope is limited to issues found by the maintainers, which may not cover all potential problems.

Provenance

Source

zai-org on Hugging Face.

Collection Method

Modification and verification of the original Terminal-Bench 2.0 dataset.

Freshness

Last updated on 2026-02-27.