OnAnOrange/qwen3.5-4b-lite-osworld-sft

Qwen3.5-4B Lite.OSWorld SFT

This checkpoint was supervised fine-tuned from Qwen/Qwen3.5-4B on the CUA-Lite Lite.OSWorld rollout data.

Training

• —Data: cua-lite/Lite.OSWorld
• —Exported SFT parquet rows: 1940
• —Training job: Slurm 1852803
• —Training wall time: 01:23:00
• —Final training step/checkpoint: iter_179
• —Hyperparameters: 8 GPUs, TP=2, MBS=1, global batch size=32, 3 epochs, LR=5e-6

Local provenance

• —Source checkout: /mnt/weka/home/lingjie.chen/cua-lite-main-train
• —Parquet: /mnt/weka/shrd/research/cua-lite/lingjie/qwen35_4b_two_train_20260708/sft/lite.osworld/train.parquet
• —Local checkpoint: /mnt/weka/shrd/research/cua-lite/lingjie/qwen35_4b_two_train_20260708/ckpts/qwen3_5-reasoning-4b/lite.osworld/sft/iter_179