Qwen3-VL-2B-OSWorld-Distill-Phase1

This model is the result of Phase-1 off-policy distillation for OSWorld GUI automation.

Training Details

  • Base Model: Qwen/Qwen3-VL-2B-Instruct
  • Training Stage: Phase-1 Off-Policy Distillation (SFT on teacher trajectories)
  • Checkpoint: iter_0000049 (best validation loss: 1.50)
  • Teacher Model: Qwen3-VL-32B-Instruct
  • Training Framework: THUDM/slime (FSDP backend)
  • Hardware: 4x NVIDIA H100 PCIe

Training Data

  • OSWorld GUI automation trajectories
  • Step-grounded format with screenshots and action labels
  • 305 distillation samples (teacher rollouts)

Intended Use

This is an intermediate checkpoint for OSWorld VLM pipeline development. Phase-2 on-policy distillation and GSPO training recommended for final model.

Citation

Downloads last month
15
Safetensors
Model size
2B params
Tensor type
F32
·
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for Jarrodbarnes/Qwen3-VL-2B-OSWorld-Distill-Phase1

Finetuned
(94)
this model