Researchers have developed TerminalTraj, a new pipeline designed to generate large-scale, executable, and verifiable terminal trajectories for training agentic models. This system addresses the challenge of creating diverse Dockerized environments and ensuring task verifiability. The pipeline has been used to curate 32,000 Docker images and generate over 50,000 verified trajectories across eight domains. Models trained on this data using the Qwen2.5-Coder backbone showed significant performance improvements on the TerminalBench benchmark, with TerminalTraj-32B achieving competitive results among models under 100 billion parameters. AI
IMPACT This work provides a scalable method for generating high-quality training data, potentially accelerating the development of more capable terminal-based AI agents.
RANK_REASON The cluster contains an arXiv paper detailing a new method for generating training data for AI agents. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
