<h1 align="center">TerminalTraj</h1> <div align="center"> 🤗 <a href="https://huggingface.co/m-a-p/TerminalTraj-32B"><b>Model</b></a>&nbsp&nbsp | &nbsp&nbsp 🤗 <a href="https://huggingface.co/datasets/m-a-p/TerminalTraj"><b>Data</b></a>&nbsp&nbsp | &nbsp&nbsp 🤗 <a href="https://huggingface.co/datasets/m-a-p/TerminalTraj-5k-instances/tree/main"><b>Instance-5k&nbsp&nbsp(with docker)</b></a>&nbsp&nbsp | &nbsp&nbsp 🤗 <a href="https://huggingface.co/papers/2602.01244"><b>HF&nbsp&nbspPaper</b></a>&nbsp&nbsp | &nbsp&nbsp 📑 <a href="https://arxiv.org/abs/2602.01244"><b>arXiv</b></a>&nbsp&nbsp </div>

This is the repo for the paper TerminalTraj: Large-Scale Terminal Agentic Trajectory Generation from Dockerized Environments

News

[2026.03.13] We have released 5,660 instances equipped with Docker-based environments! Check them out on TerminalTraj-5k-instances.

Brief

Training agentic models for terminal-based tasks critically depends on high-quality terminal trajectories that capture realistic long-horizon interactions across diverse domains. However, constructing such data at scale remains challenging due to two key requirements: Executability, since each instance requires a suitable and often distinct Docker environment; and Verifiability, because heterogeneous task outputs preclude unified, standardized verification.

To address these challenges, we propose TerminalTraj, a scalable pipeline that (i) filters high-quality repositories to construct Dockerized execution environments, (ii) generates Docker-aligned task instances, and (iii) synthesizes agent trajectories with executable validation code. Using TerminalTraj, we curate 32K Docker images and generate 50,733 verified terminal trajectories across eight domains.

Models trained on this data with the Qwen2.5-Coder backbone achieve consistent performance improvements on TerminalBench (TB), with gains of up to 20% on TB 1.0 and 10% on TB 2.0 over their respective backbones. Notably, TerminalTraj-32B achieves strong performance among models with fewer than 100B parameters, reaching 35.30% on TB 1.0 and 22.00% on TB 2.0, and demonstrates improved test-time scaling behavior.

Repository and License Metadata

For transparency and compliance, we provide a ./source/repo&license.jsonl file that records the source repositories associated with the currently released subset of our data, together with their corresponding license information.

The licenses of the repositories used in this release are restricted to the following categories: MIT, Apache-2.0, BSD-3-Clause, and BSD-2-Clause.

Third-party repositories and their contents remain subject to their original licenses. Users of this release are responsible for complying with the license terms of the corresponding upstream repositories.

Method

We propose TerminalTraj, a large-scale pipeline for generating Docker-aligned terminal agent trajectories from real-world GitHub repositories, with instance-specific executable validation.

To scale environments beyond heuristic repository filtering, we cast repository selection as model-based quality scoring, enabling automated construction of 32,325 Docker images across eight programming languages. We further curate instances spanning eight specialized domains with real-world tools and dependencies.

TerminalTraj filters rollouts via task-specific executable validators (inspired by TerminalBench). Overall, TerminalTraj produces 50,733 verified trajectories and supports continual, scalable data synthesis.

Results

As shown in the figure above, TerminalTraj-32B achieves state-of-the-art performance among models under 100B parameters on both TB1.0 and TB2.0, and its performance is close to Qwen3-Coder-480B.

In addition, we find that TerminalTraj, with its large-scale agentic data grounded in real-world environments, can substantially enhance a model’s test-time scaling capability.

Dataset

We have released our training dataset (trajectory):

from datasets import load_dataset

ds = load_dataset("m-a-p/TerminalTraj")

Models

We have released our TerminalTraj-7B/14B/32B models:

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_id = "m-a-p/TerminalTraj-14B"

tokenizer = AutoTokenizer.from_pretrained(model_id)

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16,   # 14B建议用fp16或bf16
    device_map="auto"            # 自动分配GPU
)

TODO

We will also release an additional 5,000 instances equipped with Docker-based environments in the near future.

Citation

BibTeX:

@misc{wu2026largescaleterminalagentictrajectory,
      title={Large-Scale Terminal Agentic Trajectory Generation from Dockerized Environments}, 
      author={Siwei Wu and Yizhi Li and Yuyang Song and Wei Zhang and Yang Wang and Riza Batista-Navarro and Xian Yang and Mingjie Tang and Bryan Dai and Jian Yang and Chenghua Lin},
      year={2026},
      eprint={2602.01244},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2602.01244}, 
}

TerminalTraj

Install / Use

README