What Is Plotted
The whole-330 comparison now uses the repaired full GEPA-best run: 330 rows, 7 failed initial rows rerun successfully, 0 remaining errors. Whole-330 bars use the same GEPA-style mean score as the optimizer formula; the metric table also shows micro precision/recall/F1 and exact match.
Whole 330 GEPA mean
0.7350
v10 0.7307, delta +0.0043
Whole 330 Micro-F1
0.8206
v10 0.8231, delta -0.0025
Exact Match
0.5424
v10 0.5242, delta +0.0182
FP / FN
110 / 116
v10 102 / 119
Split Scores With Whole-Dataset Context
v10 seedGEPA best
Whole 330 Metrics
v10 seed repairedGEPA best repaired
Score Table
| Set | Candidate | Score | Score meaning |
|---|---|---|---|
| Pareto 60 | v10 seed | 0.5742 | GEPA/Pareto score |
| Pareto 60 | GEPA best | 0.6979 | GEPA/Pareto score |
| Heldout 78 | v10 seed | 0.7686 | GEPA heldout mean score |
| Heldout 78 | GEPA best | 0.7882 | GEPA heldout mean score |
| Whole 330 | v10 seed | 0.7307 | GEPA mean score |
| Whole 330 | GEPA best | 0.7350 | GEPA mean score |
Whole Metric Table
| Metric | v10 seed repaired | GEPA best repaired | Delta |
|---|---|---|---|
| GEPA mean | 0.7307 | 0.7350 | +0.0043 |
| Precision | 0.8344 | 0.8246 | -0.0098 |
| Recall | 0.8120 | 0.8167 | +0.0047 |
| Micro-F1 | 0.8231 | 0.8206 | -0.0025 |
| Exact | 0.5242 | 0.5424 | +0.0182 |
| False positives | 102 | 110 | +8.0000 |
| False negatives | 119 | 116 | -3.0000 |