You can deploy this model and start earning money today!

wincentIsMe/
Qwen3-VL-2B-Instruct-SFT-VSI590k-32f-384px-tune-vision-tower

Name: Qwen3-VL-2B-Instruct-SFT-VSI590k-32f-384px-tune-vision-tower
Availability: InStock

Qwen/Qwen3-VL-2B-Instruct

finetune

endpoints_compatible

Views

No views yet

Qwen3-VL-2B-Instruct-SFT-VSI590k-32f-384px-tune-vision-tower

Full-parameter SFT of Qwen/Qwen3-VL-2B-Instruct on the VSI590k spatial-reasoning dataset (video inputs sampled at 32 frames, 384px resolution), with the vision tower unfrozen during training. See the variant without -tune-vision-tower for the frozen-vision-tower run.

Architecture: Qwen3VLForConditionalGeneration (model_type: qwen3_vl)
Training: 1 epoch, 2308 steps (DeepSpeed; trained 2026-04, uploaded from cvgl cluster)
Contents: final consolidated model.safetensors + tokenizer/processor configs — ready for inference or continued fine-tuning. Intermediate checkpoints and optimizer states are not included.

python

from transformers import AutoModelForImageTextToText, AutoProcessor
model = AutoModelForImageTextToText.from_pretrained("wincentIsMe/Qwen3-VL-2B-Instruct-SFT-VSI590k-32f-384px-tune-vision-tower")
processor = AutoProcessor.from_pretrained("wincentIsMe/Qwen3-VL-2B-Instruct-SFT-VSI590k-32f-384px-tune-vision-tower")