FACET, a terminal-task synthesis framework from Kou Shi and 12 co-authors at the University of Science and Technology of China, was published on Aug 19 and submitted to Hugging Face Papers on Aug 21. The paper ranked as the #2 Paper of the day and had 100 upvotes on the listing. FACET stands for Fine-grained Agentic Construction of Executable Tasks and is designed to create executable terminal tasks for training terminal agents.

The authors say terminal-agent training needs scalable executable supervision, but high-quality terminal tasks are difficult to synthesize because each task combines an instruction, an initialized environment, a reference solution, and an executable verifier. If those artifacts are built from inconsistent assumptions, the task can become unsolvable or be evaluated incorrectly. Multi-stage synthesis can also lose goals, dependencies, state transitions, and procedural constraints from the original sources.

FACET addresses those issues by preserving source intent and improving cross-artifact consistency. The framework reconstructs related agent skills into coherent scenarios, realizes and repairs the execution environment, and then generates the final task artifacts. A repaired container state (saved software runtime condition) grounds the instruction, solution, and verifier in the same executable setting.

The framework uses execution-based validation and targeted repair to fix artifact-specific failures without regenerating components that are already valid. According to the abstract, FACET produces complex terminal tasks with dense executable checks, and successful trajectories from those tasks provide data-efficient supervision. Fine-tuning models across multiple scales consistently improved performance on Terminal-Bench 2.1, while analyses of alternative generation schemes supported environment-grounded construction for task validity and solution-verifier alignment. The Hugging Face page also listed 18 GitHub links or references, 3 models citing the paper, 1 dataset citing it, 0 Spaces citing it, and 2 collections including it.

FACET, a terminal-task synthesis framework from Kou Shi and 12 co-authors at the University of Science and Technology of China, was published on Aug 19 and submitted to Hugging Face Papers on Aug 21. The paper ranked as the #2 Paper of the day and had 100 upvotes on the listing. FACET stands for Fine-grained Agentic Construction of Executable Tasks and is designed to create executable terminal tasks for training terminal agents.

The authors say terminal-agent training needs scalable executable supervision, but high-quality terminal tasks are difficult to synthesize because each task combines an instruction, an initialized environment, a reference solution, and an executable verifier. If those artifacts are built from inconsistent assumptions, the task can become unsolvable or be evaluated incorrectly. Multi-stage synthesis can also lose goals, dependencies, state transitions, and procedural constraints from the original sources.

FACET addresses those issues by preserving source intent and improving cross-artifact consistency. The framework reconstructs related agent skills into coherent scenarios, realizes and repairs the execution environment, and then generates the final task artifacts. A repaired container state (saved software runtime condition) grounds the instruction, solution, and verifier in the same executable setting.

The framework uses execution-based validation and targeted repair to fix artifact-specific failures without regenerating components that are already valid. According to the abstract, FACET produces complex terminal tasks with dense executable checks, and successful trajectories from those tasks provide data-efficient supervision. Fine-tuning models across multiple scales consistently improved performance on Terminal-Bench 2.1, while analyses of alternative generation schemes supported environment-grounded construction for task validity and solution-verifier alignment. The Hugging Face page also listed 18 GitHub links or references, 3 models citing the paper, 1 dataset citing it, 0 Spaces citing it, and 2 collections including it.