Unsloth Studio: local fine-tuning and PDF dataset setup

This sponsored tutorial walks through Unsloth Studio installation, local LLM fine-tuning and dataset creation. The presenter opens the browser interface at localhost:8888 and demonstrates model chat on a MacBook with 128 GB of RAM. He describes the software as open source and capable of offline operation, though the workflow shown downloads resources and later uses a cloud API.

For training, he selects the Unsloth version of Qwen 3.6 27B from Hugging Face in Safetensors format and keeps QLoRA enabled. The finance Alpaca dataset supplies instruction-and-answer examples. He reduces context length to 1,024, sets batch size to one and starts a 20-step demonstration. The selected model download is reported as 51 GB. Training loss falls during the run, but the tutorial does not establish improved finance performance with an evaluation. He also reports an MLX allocation bug that can cause failures on Apple silicon despite available memory, suggesting a smaller 9B model as a fallback.

The Recipes section turns an 80-page Nvidia financial report into question-and-answer records. The presenter configures OpenRouter, encounters errors with Sonnet 4.6 and completes a five-record test with Gemini 3.5 Flash before launching a 1,000-record run. He explains that a local model endpoint can replace the cloud provider when documents need to stay on the user's machine.