Qwen3.5-4B Desktop-Use SFT — Epoch 3
This checkpoint was supervised fine-tuned from Qwen/Qwen3.5-4B on combined
CUA-Lite Lite.OSWorld, Lite.ScaleCUA, and Lite.CUAGym desktop-use trajectories.
Training
- Data:
cua-lite/Lite.OSWorld + cua-lite/Lite.ScaleCUA + cua-lite/Lite.CUAGym
- Exported SFT parquet rows: 10,934 (including 5,000 sampled Lite.ScaleCUA rows)
- Training job: Slurm
1986549
- Checkpoint:
iter_1022, saved at the end of epoch 3 of a 3-epoch run
- Hyperparameters: 8 H200 GPUs, TP=2, DP=4, MBS=1, global batch size=32, LR=5e-6
- Fixed BSHD maximum sequence length: 18,432
This is a model-weight checkpoint. The training job did not save optimizer
state, so this artifact is suitable for inference or weight initialization but
not exact optimizer-state resumption.
Provenance
- CUA-Lite source commit:
8d2205b146a9857ed508f55bb09ba60d8e524b0c
- Base model:
Qwen/Qwen3.5-4B
