Player not loading? Watch on YouTube
This tutorial walks through fine-tuning Gemma 4 E4B IT in Unsloth Studio using the Bitext customer support dataset. The presenter uses the interface to configure training and compare the result with the original instruction-tuned model.
The demonstration runs on a RunPod A40 with a PyTorch template. The presenter chooses a rented GPU because their own GPU has 8 GB of memory and the model files total 16 GB. Those figures explain the choice in this example; they do not establish a minimum training requirement. Setup includes exposing port 8889 and running the Unsloth installer in a Linux terminal. The presenter also points to installation options for macOS, WSL and Windows PowerShell.
Training uses 4-bit QLoRA. The walkthrough covers dataset preview, AI-assisted prompt mapping and LoRA settings, then selects AdamW 8-bit optimization, a linear learning-rate schedule and a batch size of two. Unsloth Studio also accepts local datasets. Although the parameter walkthrough suggests 500 steps, the demonstrated result comes from just 30 steps.
The presenter exports a merged model to Hugging Face using a write token, then loads it into the chat comparison view. An order-cancellation prompt produces a response the presenter considers closer to the training examples. This single comparison illustrates a change in response style rather than establishing broader accuracy or customer-support reliability.
