Loading dataset details…
TMax 13K v1.1 — Harbor
All 13,565 tasks from TMax 13K v1.1, packaged in the Harbor task format. Instructions, verifiers, and supporting test assets are preserved byte for byte from the hardened source archive. Historical setup recipes are included in the Hugging Face archive. Every task uses the Open Instruct variant's public Docker image as its base, with a small wrapper that prepares the same runtime workspace.
Other formats: Open Instruct training data · Task details and audit metadata.
Harbor registry
Run all 13,565 tasks directly from Harbor Hub:
uvx harbor run -d "tmax/tmax-13k-v1.1@v1.1" --env docker \
-a terminus-2 -m "$MODEL"The registry preserves the instructions, verifiers, supporting test assets, and Docker wrappers. The Hugging Face archive also retains the historical setup.sh recipes for reference; the registry uses the already initialized public images and does not rerun those recipes.
Hugging Face download
Download and extract the task bundle with its executable permissions:
hf download TMaxxx/tmax-13k-v1.1-harbor tasks.zip extract_tasks.py dataset_release.json \
--repo-type dataset --local-dir tmax-13k-v1.1-harbor
python tmax-13k-v1.1-harbor/extract_tasks.py
# Set MODEL to the provider/model you want to evaluate.
uvx harbor run -p tmax-13k-v1.1-harbor/tasks --env docker \
-a terminus-2 -m "$MODEL"The Parquet train split is a task index for browsing and analysis. Harbor runs the task directories extracted from tasks.zip.
Task format
task_000000_f8baca82/
├── instruction.md
├── task.toml
├── environment/
│ ├── Dockerfile
│ └── docker-compose.yaml
└── tests/
├── test.sh
└── ...task.tomlcontains the task ID, taxonomy, audit metadata, and public base image inmetadata.docker_image. Harbor package names usetmax/<task_id>-v1.1; the original task IDs are retained inmetadata.task_idand the Parquet index.environment/Dockerfileuses the already initialized task image and prepares/workspace,/output, the log directories, and/app, matching the training harness. Harbor builds this small wrapper rather than bypassing workspace setup with a prebuilt-image shortcut.setup.shin the Hugging Face archive is the original build recipe, retained for provenance; it is not rerun because doing so could undo image hardening. The Compose override selects Linux/AMD64, including when using Docker emulation on ARM hosts.tests/includes the exact hardened verifier and any hidden truth/oracle assets. Harbor uploads these files at verification time; the environment build context contains no tests or setup recipes.- The agent timeout is 600 seconds and the verifier timeout is 300 seconds. These runtime limits can be overridden in Harbor. There are no reference agent solutions in this export.
The image mix is unchanged: 11,852 original images and 1,713 hardened images. See the training variant for image manifests and provenance.
Validation and provenance
dataset_release.json pins the source revisions and archive checksum. Preparation validates all 13,565 tasks with Harbor, reconciles IDs against both companion datasets, and checks that all 41,302 source files remain byte-identical in the task bundle. validation.json includes the packaging checks and representative Docker/Harbor smoke results. Runtime smoke checks validate packaging and reward collection; they do not establish full-corpus solvability.
TMax v1.1 is the hardened, prefiltered corpus, including tasks without a demonstrated passing solution. Known verifier defects and validation limitations are preserved in AUDIT.md. The paper describes the original generation recipe.
License
Released under ODC-BY, following the original TMax dataset. Research and educational use should follow Ai2's Responsible Use Guidelines. Gemini 3.1 Pro outputs included in the data remain subject to Google's Terms of Service.
Citation
@misc{ivison2026tmaxsimplerecipeterminal,
title={Tmax: A simple recipe for terminal agents},
author={Hamish Ivison and Junjie Oscar Yin and Rulin Shao and Teng Xiao and Nathan Lambert and Hannaneh Hajishirzi},
year={2026},
eprint={2606.23321},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2606.23321},
}