---
license: other
license_name: upstream-and-component-licenses
license_link: https://huggingface.co/datasets/zhangdw/TB21/blob/main/NOTICE.md
pretty_name: TB21 — Terminal-Bench 2.1 tasks and Docker images
tags:
  - terminal-bench
  - benchmark
  - docker
  - agents
language:
- en
size_categories:
- n<1K
---

![TB21 — Terminal-Bench 2.1 task and image archive](assets/banner.svg)

**TB21** packages Terminal-Bench **2.1** tasks, their Linux amd64 Docker images, and a streaming restore script for agent evaluation. Each repository contains its own task files and image archives and can be used independently.

<table width="100%" style="display: table; width: 100%; table-layout: fixed;">
<tr>
<td width="33%" align="center" style="text-align: center; padding: 16px 8px;"><strong style="font-size: 24px;">89</strong><br>Tasks</td>
<td width="33%" align="center" style="text-align: center; padding: 16px 8px;"><strong style="font-size: 24px;">89</strong><br>Docker images</td>
<td width="33%" align="center" style="text-align: center; padding: 16px 8px;"><strong style="font-size: 24px;">19.43 GiB</strong><br>Compressed images</td>
</tr>
</table>

<p align="center">
<a href="#quick-start">Quick start</a> · <a href="#whats-inside">What's inside</a> · <a href="#provenance">Provenance</a>
</p>

## Quick start

Requires **uv, Python 3.11+, Docker, and zstd**.

```bash
uvx hf download zhangdw/TB21 --type dataset --local-dir TB21
python3 TB21/load_images.py TB21
```

Use **`TB21/data/`** as the local dataset directory in [Harbor](https://github.com/harbor-framework/harbor).

<details>
<summary><strong>Load selected tasks or verify the archives</strong></summary>

```bash
python3 TB21/load_images.py TB21 --task log-summary-date-ranges
python3 TB21/load_images.py TB21 --verify-only
```

Repeat `--task` to select more tasks; all images required by the selected tasks are included. `--verify-only` checks archive sizes and SHA-256 without Docker.

The loader streams ordered parts directly into `zstd` and `docker load`. Reserve space for both the downloaded archives and Docker's extracted images. The loader restores task-file executable modes automatically.

Images target Linux amd64; ARM hosts need emulation. Agent installation and tasks that use external services may still require network access.

</details>

## What's inside

| Path | Contents |
| --- | --- |
| `data/<task>/` | Original instructions, environment definitions, tests, and solutions |
| `images/` | zstd-compressed Docker archives, split into parts of at most 1 GiB |
| `images/manifest.json` | Image references, digests, ordered parts, and SHA-256 |
| `load_images.py` | Archive verification and image restoration |

## Provenance

Official task selection: `terminal-bench/terminal-bench-2-1@6`. All 89 task-package hashes were verified against that snapshot. Task bytes, author notices, and benchmark canaries are preserved.

[Source revisions](provenance.json) · [File checksums](data-manifest.json) · [Task-package hashes](task-manifest.json)

<details>
<summary><strong>Verification scope</strong></summary>

Packaging checks covered task hashes, image layers, archive integrity, and uploaded files. Native Docker loading, model inference, and benchmark execution were not run during packaging.

</details>

Task sources use Apache-2.0; bundled components retain their own licenses. See [NOTICE](NOTICE.md) and [the task license](LICENSE-TERMINAL-BENCH). This is a community archive for evaluation; keep benchmark evaluation data out of training corpora.

**Related datasets:** [TB2](https://huggingface.co/datasets/zhangdw/TB2) · [TB4](https://huggingface.co/datasets/zhangdw/TB4) · [GAIA2-mini](https://huggingface.co/datasets/zhangdw/GAIA2-mini)