Is this actually working?
#10
by
Weyaxi
- opened
Hi @fblgit ,
Nice idea to create a dataset like this :)
I wonder, can this dataset actually improve the performance of the model? Have you tested anything?
Yeah, MATHQA 7 whole points on PHI-2 :D from 30 to 37+ also boosted 2 points on ARC, but lost 2 in GSM (i guess formatting, complexity).
But it needs to be curriculum, gradually increasing complexity. like min/max & floats && output.floats gradually increasing little by little, in chunks like 50ks~