Localpager GEPA Reports

Default dashboard for the Evalstate Qwen GEPA overlay run. Updated 2026-06-17.

Bottom Line

Full 330 GEPA mean
0.7350
v10 0.7307, delta +0.0043
Full 330 Micro-F1
0.8206
v10 0.8231, delta -0.0025
Precision / Recall
0.8246 / 0.8167
precision down, recall up
FP / FN
110 / 116
v10 102 / 119
Heldout Micro-F1
0.8417
v10 0.8296, delta +0.0121
Best Pareto score
0.6979
seed 0.5742

GEPA-best is not a clear replacement for v10. It improves the GEPA objective and exact match, but false positives increase and full-set micro-F1 is slightly lower.

Whole 330 Score Graph

0.500.600.700.80 v10 GEPA mean 0.73070.7307GEPA best GEPA mean 0.73500.7350GEPA mean v10 F1 0.82310.8231GEPA best F1 0.82060.8206Micro-F1 v10 precision 0.83440.8344GEPA best precision 0.82460.8246Precision v10 recall 0.81200.8120GEPA best recall 0.81670.8167Recall full 330 repaired outputs
v10 seedGEPA best

Open The Detailed Graphs

Proposal GraphsEvery proposal attempt, accepted/rejected status, subsample deltas, and best-so-far Pareto score. Whole Dataset ComparisonFull 330-row GEPA-best versus v10 charts and metric table. Iteration GraphOriginal GEPA score report for the run. Prompt Diff PickerDropdown-to-dropdown comparison for individual candidate prompts. Candidate TreeGEPA-native candidate tree visualization. Final ReportRun settings, heldout comparison, whole-330 check, and artifact links.

Metric Table

Metricv10 seedGEPA bestDelta
GEPA mean score0.73070.7350+0.0043
Micro-F10.82310.8206-0.0025
Precision0.83440.8246-0.0098
Recall0.81200.8167+0.0047
Exact match0.52420.5424+0.0182
False positives102110+8
False negatives119116-3

Archive

2026-06-14 12B cardinality report 2026-06-14 12B score graph 2026-06-14 prompt diffs