Unsloth Finetuning Playbook - Fine-tuning GPT-OSS-20B with GB10 Forum Data

Accelerated Computing DGX Spark / GB10 User Forum DGX Spark / GB10 Projects
, ,
1

Hi everyone in the GB10 forum!

I’ve been following the discussions here and wanted to contribute something back to this great community. In honor of all the pioneers exploring the DGX Spark system, I’ve put together a “Finetuning Playbook” using the contents of this very forum.

I created a pipeline that allows you to fine-tune the GPT-OSS-20B model directly on your DGX Spark using Unsloth.

What this playbook does:

  1. Scrapes this very forum: It downloads questions and answers from the DGX Spark GB10 category. (I’ve checked with @aniculescu before publishing to verify compliance with NVIDIA terms, so it’s ok)

  2. Creates a Dataset: It uses a local LLM (via LM Studio) to clean up the threads and create a high-quality Q&A dataset.

  3. Fine-tunes efficiently: It uses Unsloth to fine-tune GPT-OSS-20B on the Spark’s GPUs.

  4. Exports for Inference: Includes scripts to export the fine-tuned model to GGUF for use with Ollama or LM Studio.

image
image2798×1998 389 KB

It’s a fun way to test the capabilities of the Spark and create a model that is knowledgeable about the specific issues and solutions we discuss here!

You can check out the code and instructions here:

I hope you find this useful for your own projects. Let me know if you have any questions or run into any issues!

Cheers,

Raphael

2

It still didn’t resist hallucinating 8-port Spark :)

3

BTW, you can delegate post cleaning (slang, signature) to LLM too.

I’ve had much better luck fine-tuning non-thinking models, though, especially with limited datasets.