Usage notes
Uses the same chat template as Qwen3-VL-4B-Thinking; serve it like the released OpenWebRL-4B-SFT checkpoint.
Training
- Data: OpenWebRL SFT trajectories, 3,085 per-turn rows, rendered with LlamaFactory's
qwen3_vl template (history masked; only the final assistant turn of each row is trained).
- Full fine-tune with LlamaFactory: vision tower frozen (381,783,040 params); projector and language model trained (1,745,748,992 of 2,127,532,032 params).
- 3 epochs (75 optimizer steps), learning rate 1e-5, cosine schedule, warmup 0.1, effective batch 128 (1 per device x 64 accumulation x 2 GPUs), bf16, DeepSpeed ZeRO-2, cutoff 36,864 tokens, image_max_pixels 12,845,056.
- 2 x H200, about 45 minutes.
- Mean training loss by epoch: 0.801 / 0.608 / 0.531. No held-out split was used; most of the drop after epoch 1 comes at epoch boundaries (repeat passes over the data).
Evaluation
Not yet evaluated.