Open AG161 opened 2 days ago
An alternative would be upgrading these repos to the latest version of Lean. Is that feasible?
That would be possible in most cases but not ideal because it would not be a direct apples-to-apples comparison. For example in the miniF2F repo I linked, mathd_algebra_176
in Test.lean cannot be closed by ring
and requires a more complicated proof because the exponent is a real number and not a natural number, in the current version it can be closed just by ring
because the functionality of that tactic was expanded. Also I have not tried updating the Lean version of an older version of mathlib4, but it's possible that for a project of that size it would not be trivial - I'm not sure though since I don't have any experience with that.
You could edit the installed *.py directly to remove traced_repo.check_sanity()
.
I know that LeanDojo has dropped support for older versions of Lean, but I used to be able to trace older Lean repos with older LeanDojo versions, now I'm getting errors when I do so.
Here are two examples with the following miniF2F repo: https://github.com/rah4927/lean-dojo-mew/
LeanDojo 1.3.0 logs (I removed the 1k+ lines of the repo's building):
LeanDojo 1.5.1 logs (again with build lines removed):
I understand support for older Lean versions has been dropped. Still if possible I would appreciate some insight in case the problem is simple, because many of the machine learning benchmarks such as miniF2F are on older versions of Lean, and if I cannot trace the repos I cannot evaluate on these benchmarks. Thanks!
Platform Information