Unsloth Studio is insane… fine-tune any AI model locally

Unsloth studio: https://unsloth.ai/docs/new/studio Get everything from the video: https://www.davidondrej.com/local-ai-finetuning-starter-kit

The video is not complete , I didn't see how to use the fine-tuned model , the delivery of the entire workflow should be a new weights file (or files) , that you can later use in any LLM server , like vLLM and alike

Great video! My one queston, is whether unsloth studio can create a combined dataset training bundle that can handle personalisation at the same time as PDF fine-tuning ? ie., I have dialogue transcripts of the Simpsons + NVidia financials PDF. I want the dataset recipe to combine dialogue + PDF by the deepseek-v4 distillation. Ie., ultimately, I want to fine-tune Qwen 3.6 model with NVidia financial year-ends, to respond in the personality of Krusty the Clown ? (fake example of course)

I want to know how use lora in agents ai because full fine-tune takes alot of time , can you please make video how to use lora adaptor in agent ai

Great tutorial to get started, but it feels like it only scratches the surface of the real issues we face during fine-tuning. I'm currently running into severe model collapse (mode collapse/repetition) and overfitting, which weren't covered here. In my recent runs, no matter how much I aggressively tune down the learning rate, epochs, or rank, the model either underfits completely or collapses into a degenerative state (e.g., repeating the same outputs or ignoring the prompt context entirely). It feels like it's just finding a shortcut to minimize loss by memorizing formats rather than learning the underlying distribution. Is anyone else experiencing this exact same degradation? And I am now working on using dora.

Can this be done for voice agent? Have you ever created one as an outbound agent, or any way you can show that in a video?

Dude... You rock.

What if you trained your LLM, then the same LLM that you trained got a huge upgrade from the source? You will you be, dame what i have now is obsolete and i made efforth to train it? Well i am just starting to learn, would this be possible? What should be done?

Great Video - Thank You

Thank you for this one!