Using SRBench in 2026

[fredericosantos]

Hello,

I am struggling to follow how to best apply SRBench in 2026. Did I understand correctly that first principles would be scored through the expression itself? --sym_data flag?

What branch should I be using? I'm trying to use SRBench for pure GPU computation and any cpu-syncing would severely hurt the algorithm's performance.

Is there a definite 2026 guide? Do I have to run it through the docker branch? Or the main branch? Or the 2025 branch?

PS: as a side note, I would like to point out that the vineyard dataset contains 2 features and 53 rows, and it was used as a negative example for regression (when not to use it). Algorithms are not able to optimize it because there is too little information. Would the team be open to discussion into which datasets make sense to be included? I would love to contribute with more up-to-date datasets (vineyard is from 1990 for example) that push the limits of algorithms.

[folivetti]

Hi @fredericosantos , I'll let @gAldeia answer the more practical questions as he is more involved in getting the new version up and running. Let me just tell you in advance that he is working on a major update to make things easier for the organizers and the users!

Regarding datasets, yes, this is currently a major issue we are having. Please, join the dicussion here #174 thank you

[fredericosantos]

thank you. I would also like to understand if the 1h limit is per run or per 30 runs.

and if the 6h of hpo are per dataset, or in total

[fredericosantos]

A follow-up question: is the 6 hours of possible hyperparameter optimization the paper mentions per seed, per dataset, or per entire SRBench?