For anyone interested in fine-tuning here’s a great beginner friendly guide on the basics and best-practices.
6 Likes
Thanks, Justin!
Fine-tuning will absolutely help to build the next generation of models. At even smaller size.
What was my interesting experience? For one of my use cases, it was easier to rebuild the model.
The complete model build took one hour on a Mac Studio, and it completed within a few hours on my MacBook Air.
The news sells model building so complex and resource-hungry, so most people never start such a journey.
@hello1 would love to get any insight into your approaches for full model building - your right that getting started seems v scary but its something that if we get more people doing will help all - so yeah any links or docs you have about how to start would be great to see
I use AI for Vibe coding for these codes.
The first step is to build question and answer pairs.
For the model, I asked my used model to implement a teaching with the HRM method.
With 100K samples it created fairly good results.
@hello1 Do you have any more details on how you did it? I see fine tuning models (particarly gpt-oss and now deekseep v3.1) as being an important step - and doing it on a Mac Studio quickly is very cool..
I collected information in markdown format.
Then I wrote a program that creates factual statements and questions based on the markdown content. (Chunking also useful such as for RAG).
Then I used a different large model to evaluate the question answer pairs, whether they are true or not.
Finally, I used transformers and PEFT to train the model, then evaluated it using a different set.
This is a generic training that I usually use. Now most of my research time is focused on homomorphic encryption using inference.
I found some breakthroughs, but I still need some tuning.