I need a kick start, please 🙂

[TonyGravagno]

I'm a mature developer with decades of experience building business applications and utilities. I understand the GPT model, embedding, context, etc.. I'm very competent with using modern LLMs, prompting for desired outcomes, and developing agentic flows and other applications with common APIs for AI. In summary, I'm not a kid, now new at this, and not likely to waste anyone's time.

But I've never trained or fine tuned a model. And I don't yet understand the process of training a new model like gpt-oss-safeguard with policy and examples.

I do have a couple applications, for trust policies, not safety. I'd be willing to share competent models developed.

Where can I go to read and learn the process? I'm not looking for a fish, I want to learn how to fish "here"... and I'm guessing that should make sense to anyone who can answer the question.

Let's say I download the 20b model. What's next? Where can I find a HowTo on what's expected, document formats, procedure for reward or failure resolution?

Might this be a topic for a Discord gathering? Videos? Articles?

Thanks in advance for kind guidance which I'll be happy to share with others when I can, in the spirit of OSS.

[julietshen]

Hi @TonyGravagno! I will answer as someone who's pretty new to this, and share what helped me. Others like @vinaysrao1 may have more specific resources.

• Our friends at Hugging Face have just released a super detailed guide on training models: https://huggingface.co/spaces/HuggingFaceTB/smol-training-playbook
• The user guide for gpt-oss-safeguard can be helpful for practical instructions on how to prompt the model for the best results: https://cookbook.openai.com/articles/gpt-oss-safeguard-guide#introduction--overview
• since gpt-oss-safeguard is part of the gpt-oss family, OpenAI has other documentation related to how to fine tune it, how to handle the chain of thought/reasoning: https://cookbook.openai.com/topic/gpt-oss

I love this as a topic for the first RMC office hours! We'll be announcing more details on that shortly. Let me know if this ISN'T what you're asking about, and I can try to find a better answer or ask someone else!

[TonyGravagno]

I've opened the cited references and will get through them as quickly as I can. I'll be happy to serve as an example of what we get when someone just follows the guides, reads the material, and tries to execute on the information available. If you could defer an office hours for about a week until I can confirm that I've been through all of this, then I can come to the session prepared for comment. If that doesn't fit your schedule I'll try to expedite. Thanks!!