WIP quadrupling the time horizon of the fast evals so we can get better estimates - #568
Open
eugenevinitsky wants to merge 1 commit into
Open
WIP quadrupling the time horizon of the fast evals so we can get better estimates#568eugenevinitsky wants to merge 1 commit into
eugenevinitsky wants to merge 1 commit into
Conversation
carla_fast runs 250 scenarios at scenario_length=500, which is short enough that per-scenario metrics carry a lot of variance. Quadrupling the horizon gives each scenario more steps to average over, at 4x the eval cost. Still well under the full carla benchmark's 6000. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
carla_fastruns 250 scenarios atscenario_length: 500. This raises that to2000— 4x the horizon, same scenario count.Why
500 steps is short enough that per-scenario metrics carry substantial variance, so the fast eval is a noisy estimator of the thing we actually care about. Giving each scenario 4x the steps to average over tightens the estimate without touching scenario sampling or seeds.
2000 is still well under the full
carlabenchmark's 6000, socarla_faststays meaningfully cheaper than the full pass.Notes
carla_fastrun. That hits training-time eval hardest, which fires everytrain.evaluation_interval_epochs(100 in the nightly recipe) — worth measuring the wall-clock impact on a nightly before this lands.eval_carla_fast/*across this change will show a step discontinuity that is an artifact of the horizon, not a policy change.