AuroraAI-Research/Aurora-2

I introduce Aurora-2, as the name suggests, its just 2 parameters.
The model also didnt undergo any training at all.
the method used for it, was a technique I named "Noisemaxxing".

What is Noisemaxxing?
Noisemaxxing is when you take a large number of randomly initalized models, with different seeds, and dont train them at all, just evalute them.

I used 10000 Models randomly initalized models for this, here are the statistics: INT Index: mean=2.3823 std=0.4363 min=1.0253 max=3.6442

here is the benchmarks of the released model;
Hellaswag : 23.98%
ARC-Easy : 26.73%
ARC-Challenge : 24.83%
Combined ARC : 25.78%
PIQA : 50.65%
ArithMark-3 : 39.20%

With a final Int index of 3.64.
Clearly Arithmark-3 is doing the heavy lifting

Update! A second version has been uploaded, with better scores achived by scaling down the maximum magnitude of the floating point numbers.
The new benchmark scores:

HellaSwag : 23.78%
ARC-Easy : 26.77%
ARC-Challenge : 24.66%
Combined ARC : 25.71%
PIQA : 52.94%
ArithMark-3 : 39.20%
INT Index : 4.796