Context

While using the Qwen3-VL (8B) Vision notebook to fine-tune Qwen3-VL(8B) on ChartQA, I ran into two issues. I figured I'd
contribute the fixes back.

Reference notebook: https://colab.research.google.com/github/unslothai/notebooks/blob/main/nb/Qwen3_VL_(8B)-Vision.ipynb

Issues Fixed

1. Deprecated tokenizer argument in SFTTrainer

tokenizer = tokenizer is deprecated in newer versions of TRL. The correct argument is processing_class = tokenizer, which properly handles vision processors like Qwen3VLProcessor.

2. Missing eos_token in SFTConfig

Without explicitly setting eos_token, TRL falls back to a placeholder <EOS_TOKEN> which doesn't exist in Qwen3VLProcessor's vocabulary, causing this error at runtime:

ValueError: The specified eos_token ('<EOS_TOKEN>') is not found in the vocabulary of the given processing_class (Qwen3VLProcessor). Ensure that the eos_token exists in the vocabulary before using it as an EOS token.

Fix: pass eos_token = tokenizer.tokenizer.eos_token in SFTConfig to use the inner tokenizer's EOS token.

Changes

• original_template/Qwen3_VL_(8B)-Vision.ipynb - source fix
• Regenerated nb/ and python_scripts/ via update_all_notebooks.py

Test Plan

• Ran the notebook end-to-end with 4-bit quantization - SFTTrainer initializes and trains without error