I used to only know how to use chatgpt and gemini, and I never really thought about tuning them.
While working on a project, I wanted to introduce AI, but security concerns prevented me from uploading data to web-based AIs.
I only thought about LLMs and did a very simple installation. Just checked if it was working or not.
When I checked the workstation I was using, it had an RTX 2000 ADA or an RTX PRO 2000 blackwell.
Both are 16G models, but they are far from sufficient for running AI seriously.
Anyway, I downloaded the qwen3.5-4B model to test on my workstation and
Created a very simple example generator and trained it through that example.
Even though it's a 4B model, training uses more than 9G of memory, which is a bit disappointing due to hardware limitations.
Before training, after training 100, 1000, and 10000 examples, I gave the same very simple problem to each and asked for a solution.
Before training: It thought about something but couldn't find an answer within the limited tokens.
100: The way of thinking changed, and it produced some results, but the answers were wrong.
1000: Thinking became much simpler than 100 cases, and it found the correct answer.
10000: Thinking was a bit longer than 1000 cases, but it correctly identified the cause and found the right answer.
It's still at a very basic level, far from practical use, but it's amazing to see AI learn and solve problems for the first time.
Currently, I have only confirmed the possibility of 4 stages of training, and I think it would be good to check the feasibility in more detail by proceeding to about stage 1.
I'm curious about how much a 30B, 70B, or 120B model would be... Eventually, I bought an ASUS Ascent GX10 (DGX SPARK) yesterday... (Suddenly, a 1Tb model came up for 3560 euros...)
I'm thinking about developing this result further and requesting to use the A100 cluster in my workplace.
