FACET-Terminal

FACET-Terminal is the official Hugging Face organization for FACET: Fine-grained Agentic Construction of Executable Tasks.

FACET is a framework for synthesizing complex, coherent, and verifiable terminal tasks from reusable agent skills. It is built around two key principles:

• Preserving source intent across multi-stage task synthesis.
• Grounding all task artifacts in a shared executable state, including the instruction, environment, reference solution, and verifier.

FACET reconstructs related skills into information-rich scenarios, builds and repairs the execution environment before generating final task artifacts, and uses execution-based validation with targeted repair to improve task reliability.

Starting from 71K+ agent skills, FACET constructs 6K+ execution-validated terminal tasks. Successful task trajectories provide data-efficient supervision and improve models across multiple scales on Terminal-Bench 2.1.

Resources

• 🌐 Project Website
• 💻 GitHub Repository
• 📄 Paper, datasets, trajectories, and model checkpoints — coming soon

Contact

For questions and collaboration, contact Kou Shi at stokou@mail.ustc.edu.cn.