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

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.

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.

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.

I’m currently on vacation and traveling. I’ll organize the data and models as soon as possible, and release them once the review has been approved ...