Skip to content

fix(evals): show error rate on partially evaluated epochs - #3462

Merged
samsja merged 1 commit into
mainfrom
feat/evals-log-error-rate
Sep 2, 2026
Merged

samsja merged 1 commit into
mainfrom
feat/evals-log-error-rate

Conversation

@samsja

@samsja samsja commented Sep 1, 2026 •

Copy link
Copy Markdown
Member

Summary

  • the partially-evaluated (cancelled-epoch) warning line in the uv run evals entrypoint log now shows the Error rate next to Reward — previously it reported reward with no signal of how many completed episodes errored
Partially evaluated swebench-pro (Step 240) |  1m 32s | Reward 0.4167 | Error 5.0% | Completed 92/100 | Cancelled 8/100

The completed-epoch success line already shows Error and is unchanged:

Evaluated swebench-pro (Step 240) |  1m 32s | Reward 0.4167 | Turns 3.2 | Branches 1.0 | Error 5.0% | Truncation 2.0%

Verification

  • uv run ruff check src/prime_rl/evals/evals.py — passed
  • uv run ruff format --check src/prime_rl/evals/evals.py — passed

Note

Low Risk
Logging-only change in eval finalization; no behavior, scheduling, or metric emission changes.

Overview
When an eval epoch is cancelled (e.g. superseded by a newer checkpoint), the warning log line now includes the Error rate from completed episodes, using the same full.has_error formatting as the full Evaluated success line.

Previously that partial-eval message only showed reward and completion/cancel counts, so operators could not see how many finished rollouts failed without digging into metrics.

Reviewed by Cursor Bugbot for commit 010a7ec. Bugbot is set up for automated code reviews on this repo. Configure here.

@samsja
samsja force-pushed the feat/evals-log-error-rate branch from 538a455 to eb91658 Compare September 1, 2026 21:42
@samsja samsja changed the title feat(evals): show error rate and pass@k in the eval log fix(evals): show error rate on partially evaluated epochs Sep 1, 2026
@samsja
samsja marked this pull request as ready for review September 2, 2026 03:30
The cancelled-epoch warning line reported Reward with no signal of how
many of the completed episodes errored. Show the same Error rate as the
completed-epoch success line.
@samsja
samsja force-pushed the feat/evals-log-error-rate branch from eb91658 to 010a7ec Compare September 2, 2026 20:09
@samsja
samsja merged commit c84068c into main Sep 2, 2026
18 of 19 checks passed
@samsja
samsja deleted the feat/evals-log-error-rate branch September 2, 2026 20:26
mikasenghaas added a commit that referenced this pull request Sep 14, 2026
Main moved the prime monitor onto prime-runs (#3459) and renamed the
admin clients to AdminPlane (#3472). The eval monitor now opens one
prime-runs eval run per finished epoch instead of posting to
/evaluations by hand, which retires prime_rl.monitors.platform: the
SDK's projection and metrics modules own the sample format and the run
aggregates. The eval runner takes the AdminPlane rename and the
error-rate line from #3462. The swe example stays under examples/eval
(main had moved it to configs/evals for `uv run evals`), and the
verifiers submodule follows feat/remove-eval-cli's merge of main.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants