OnAnOrange/qwen3.5-9b-lite-osworld-cuagym-sft

Qwen3.5-9B Lite.OSWorld + Lite.CUAGym SFT

This model is a supervised fine-tuning checkpoint of `Qwen/Qwen3.5-9B` for computer-use agent trajectories.

Training data

• —cua-lite/Lite.OSWorld-v1 at revision 0c5d2495f35a27eae7baf8aa50b2e677b209de39: 1,937 examples
• —cua-lite/Lite.CUAGym-v1 at revision 92fff449649d210bdfca09b71d4094900f8d5147: 1,788 examples
• —Combined total: 3,725 examples

The data was exported with examples/lite/v1/configs/qwen3_5/reasoning/lite.osworld.yaml.

Training configuration

• —Epochs: 3
• —Global batch size: 32
• —Micro batch size: 1
• —Tensor parallel size: 2
• —Data parallel size: 4
• —Learning rate: 5e-6
• —Minimum learning rate: 1e-6
• —Final training step: 347
• —Final training loss: 0.235178

Only the final checkpoint was retained. The weights in this repository are the final Hugging Face export from Slurm job 1875082 (eval-bbq-1).