> gpqa diamond alone makes up 20% of their "capability index" and is responsible for ~70% of the gap between their model and nemotron nano (10 point difference!)
> so what about contamination? i looked at the mixture and they train on a dataset "AIML-TUDA/QA-base" (for 10 epochs lol) that literally is a very very light rephrasing of the gpqa diamond test set
> tl;dr: they trained on 10 epochs of a "very light rephrasing" of EVERY gpqa diamond EVAL data :)
https://huggingface.co/Soofi-Project/Soofi-S-Instruct-Previe...
> gpqa diamond alone makes up 20% of their "capability index" and is responsible for ~70% of the gap between their model and nemotron nano (10 point difference!)
> so what about contamination? i looked at the mixture and they train on a dataset "AIML-TUDA/QA-base" (for 10 epochs lol) that literally is a very very light rephrasing of the gpqa diamond test set
> tl;dr: they trained on 10 epochs of a "very light rephrasing" of EVERY gpqa diamond EVAL data :)