ArtysicistZ/UI-TARS-MCTS-SFT-Data

--- configs: - config_name: default data_files: - split: train path: mcts_success.jsonl --- # UI-TARS MCTS SFT Training Data - `mcts_success.jsonl` — 1,852 trajectories with base64 screenshots (2.9GB) - `step_masks_train.json` — per-step loss masks (15,920 KEEP / 2,181 REMOVE) ## Loading ```python import json # Load trajectories with open("mcts_success.jsonl") as f: trajectories = [json.loads(line) for line in f if line.strip()] # Load masks with open("step_masks_train.json") as f: masks = json.load(f) ```