Now that the functionality we need is available in relevant releases, upgrading to the next release after the pinned commits. We need to use released versions of the packages to build a wheel too. Test results seem to remain unchanged (except in third decimal places, but generally an improvement where there is a minor difference in lm-eval results. lm-eval upgrade brings in a fix to stop sequence handling, where generations were sometimes cut off too early.
Now that the functionality we need is available in relevant releases, upgrading to the next release after the pinned commits. We need to use released versions of the packages to build a wheel too. Test results seem to remain unchanged (except in third decimal places, but generally an improvement where there is a minor difference in lm-eval results. lm-eval upgrade brings in a fix to stop sequence handling, where generations were sometimes cut off too early.