[Bug] Can't train Qwen3.5 or Gemma4 on multimodal datasets in Unsloth Studio

[virusapex]

Note: Please do not remove the questions. Answer beside them.

1. Did you update? pip install --upgrade unsloth unsloth_zoo
   I am running a Docker image with tag 2026.4.2-pt2.9.0-vllm-0.16.0-cu12.8-studio-release-v0.1.35-beta
2. Colab or Kaggle or local / cloud
   Local
3. Number GPUs used, use nvidia-smi
   A single H100 GPU
4. Which notebook? Please link!
   Unsloth Studio
5. Which Unsloth version, TRL version, transformers version, PyTorch version?
   Unsloth 2026.4.2, TRL 0.23.1, Transformers 4.57.1, PyTorch 2.9.1+cu128
6. Which trainer? SFTTrainer, GRPOTrainer etc
   I assume it is SFT by default, but there is no visualization in Studio.

I tried my custom .parquet dataset with images and text columns, as well as unsloth/LaTeX_OCR example, but both of them produce the same Text model is not compatible with a multimodal dataset. Switch to a vision model or choose a text-only dataset. on unsloth/Qwen3.5-4B and unsloth/gemma-4-E2B-it even though they both have vision capabilities. However, unsloth/Qwen3-VL-4B-Instruct worked fine and was successfully trained on the dataset.

[ZeroGold]

@virusapex Did you try the models that aren't working with different datasets? I see you tried a different model on the same dataset, but what are the findings when you try the same models on a different dataset?

[virusapex]

@virusapex Did you try the models that aren't working with different datasets? I see you tried a different model on the same dataset, but what are the findings when you try the same models on a different dataset?

As I mentioned, I tried a couple of combinations to which I added some more just to test the issue:

• unsloth/Qwen3.5-4B and unsloth/gemma-4-E2B-it on:
  • MMMU/MMMU - throws the error (also a weird issue where you can not find this dataset if Qwen3.5 is selected but with older models like Gemma3 or Qwen3-VL it is visible in the search field for datasets) ;
  • unsloth/LaTeX_OCR - throws the error;
  • unsloth/alpaca-cleaned - no issue;
  • unsloth/OpenMathReasoning - no issue.
• unsloth/Qwen3-VL-4B-Instruct and unsloth/gemma-3-4b-it deliver no issues when selecting above datasets.

[Datta0]

Interesting, lemme check

[rolandtannous]

@virusapex @ZeroGold please manually pull the latest image.

docker pull unsloth/unsloth:latest

The image you are using is now outdated. Thank you

[virusapex]

@virusapex @ZeroGold please manually pull the latest image.

docker pull unsloth/unsloth:latest

The image you are using is now outdated. Thank you

The latest image 734ff6abe809 has the same issue. I assume, after merging the pull request #4868, it will be applied to the Docker image.

[rolandtannous]

we don't use vllm in unsloth studio though i think you meant #4855
I might make a container version with the two patches before they're merged though

[virusapex]

we don't use vllm in unsloth studio though i think you meant #4855 I might make a container version with the two patches before they're merged though

Ok, got you. I will wait until it is merged into a Docker image.