TerminalHorizon-Environment

Scaling Agentic Data for Long-Horizon Terminal Intelligence

Paper · Project Page · Citation

TerminalHorizon is a fully automated data synthesis engine for long-horizon terminal agents. Grounded in real-world professional work, it constructs executable environments and generates training trajectories that connect agentic behavioral patterns across stages toward a shared goal.

This repository releases the 1,535 task environments underlying the project. Each task includes a public instruction, a containerized runtime, a reference solution, and a verifier in Harbor format. The companion TerminalHorizon-3K dataset provides the training trajectories.

Resources

Resource	Description
TerminalHorizon-Environment	Executable terminal-task environments.
TerminalHorizon-3K	Long-horizon training trajectories.
TerminalHorizon-27B	Fine-tuned Qwen3.5-27B.
TerminalHorizon-35B-A3B	Fine-tuned Qwen3.5-35B-A3B.
TerminalHorizon-122B-A10B	Fine-tuned Qwen3.5-122B-A10B.

Dataset Overview

Property	Value
Task environments	1,535
Professional domains	25
Subdomains	463
Environment format	Harbor task packages

The terminal is a shared interaction interface, not a restriction to software tasks. The environments span software engineering, machine learning, education, agriculture, healthcare, finance, scientific research, and other professional domains. They are designed to support substantial, sustained terminal work whose difficulty follows from the professional objective and the work needed to complete it.

TerminalHorizon uses challenge-guided reconstruction and progressive horizon expansion to develop environments for long-horizon problem solving. Its focus is on meaningful work and connected agent behavior, rather than task length alone.

Using the Environments

Download the dataset and extract the task packages:

pip install -U huggingface_hub
hf download shunzou05/TerminalHorizon-Environment \
  --repo-type dataset --local-dir TerminalHorizon-Environment

cd TerminalHorizon-Environment
mkdir -p tasks
for archive in data/tasks-*.tar.gz; do
  tar -xzf "$archive" -C tasks
done

Each task is stored intact in one archive and follows the Harbor package layout:

<task_id>/
  instruction.md       # Public task description
  task.toml            # Task and runtime configuration
  environment/         # Environment definition and task materials
  solution/            # Reference solution; entry point: solve.sh
  tests/               # Verifier; entry point: test.sh

Read the Task Index

The Parquet index provides one searchable row per task. It supports browsing instructions, categories, package summaries and Docker environment definitions without extracting the task archives.

from datasets import load_dataset

tasks = load_dataset(
    "shunzou05/TerminalHorizon-Environment",
    split="train",
)
print(tasks[0]["task_id"])
print(tasks[0]["instruction"])

Column	Meaning
task_id	Task directory name
category, subcategory	English primary and secondary categories
instruction	Complete public instruction text
dockerfile	Primary Dockerfile text
package_bytes	Sum of regular-file byte lengths before compression
file_count	Number of regular files in the task

The train split is an indexing convention, not a prescribed train/test split. The Parquet index is not a replacement for the complete task archives: fixtures, repository history, binary assets and executable task files remain in data/.

Citation

@misc{zou2026scalingagenticdatalonghorizon,
      title={Scaling Agentic Data for Long-Horizon Terminal Intelligence},
      author={Shun Zou and Ziyu Ma and Yi Zou and Yong Wang and Lin Chen and Zehui Chen and Guanhua Chen and Xiangxiang Chu and Feng Zhao},
      year={2026},
      eprint={2610.13166},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2610.13166},
}

License

The dataset is distributed under Apache-2.0. Bundled third-party code and data remain subject to their respective licenses and notices.