4.9 % reached on ARC-AGI2 benchmark

[avinavsahoo]

Hi all,

We have been training a light weight brain influenced model to benchmark on ARC- AGI-2, we have reached 4.9% on that benchmark while models like Gemini 2.5 pro are at 1.3% etc we expect to reach 50% within a month or two making it position one on leadership board. Our EM is in fraction 0.049 so percentage comes to 4.9 when u multiply with 100

[alavia]

Try increasing your epochs number to 50, maybe? I am not an expert in this training, but I have trained a couple of CNN and segmentation models, and epoch tuning always works.