ProCUA-SFT

ProCUA-SFT ProCUA-SFT is a large-scale synthetic trajectory dataset for training computer-use agents (CUAs): models that operate graphical desktop environments from screenshots using mouse, keyboard, and code-like actions. The dataset accompanies the **ProCUA-SFT Technical Report** and is designed for supervised fine-tuning of screenshot-based desktop agents. This repository contains the raw trajectory artifacts used to construct those step-prefix SFT samples: trajectory JSON files and their corresponding screenshots. To keep the Hugging Face repository reliable for a dataset with millions of small files, trajectories are distributed as compressed tar shards under `shards/`. Dataset Summary ProCUA-SFT was produced by an automated pipeline that uses a single VLM, Kimi-K2.5, in multiple roles: goal generation, precondition verification, and trajectory rollout. The pipeline synthesizes grounded desktop tasks, verifies that task preconditions hold before rollout, executes the task in a full desktop environment, and records screenshot-action trajectories. The dataset is grounded in realistic desktop content and configurations, including: multi-application desktop configurations adapted from OSWorld; spreadsheet-grounded tasks using real-world spreadsheets from SpreadsheetBench as environment inputs; presentation-grounded tasks using permissively licensed Zenodo10K presentation files as environment inputs. Dataset Contents The current release contains **93,566 raw trajectories**: | Directory | Number of trajectories | | --- | ---: | | `part_1/` | 16,558 | | `part_2/` | 59,096 | | `part_3/` | 5,912 | | `part_4/` | 12,000 | | **Total** | **93,566** | The four `part_*` directories are storage shards for organization and distribution. They should not be interpreted as train/validation/test splits. On the Hub, the raw directories are packaged into approximately 50 `.tar.zst` files: Each archive contains paths rooted at the original `part_*` directories, for example: The accomp