Fine-tuning jobs

Good day.
I try to fine-tuning little model meta-llama/Llama-3.2-3B (LoRA) on a small dataset of 50 questions. I started Mar 23, 2025, 9:55:24 AM (job ID - ftjob-90e653a21fd34ab2a2bab6be00557764) and it is still running! (30+ mins)

It is way too long.

I think it is because you place all fine-tuning jobs in one queue.
I think it is wrong approach. You can divide jobs at least on 2 queue: for small models and for big models (by base model sizes in *B).

And to be honest - you can put more GPU on fine-tuning jobs (only for fine-tuning). 1 server that you separete ONLY for fine-tuning jobs for small models can handle all fine-tuning jobs in seconds!

Training little models with GPU that Nebius have MUST be lightning fast - it can be your main advantage.

For us → the faster we train → the faster we do our experiments → the faster we make money