---
license: cc-by-4.0
tags: [medical-imaging, intracranial-hemorrhage, rsna, variance-study]
---

# Sequence-head variance study — `ICH_ef04_A_baseline_b03-v2_seed42`

Step 6's `A_baseline` beat EF03 on validation (0.9526 vs 0.9472) and lost on test
(0.9569 vs 0.9772). This study isolates why: the 12-epoch sequence-head val curve
swings ~0.06 on a flat loss, so "best val" selects the maximum of twelve noisy draws.

Sweeps learning rate x selection rule x seed, training only the sequence heads from
frozen Step-6 embeddings (~3 min per run). The config is chosen on **val-side
stability** — seed spread and val-curve swing — never on test.

`scores/seed*/` hold Step-2-schema dumps for the chosen config, one directory per
seed, so Step 4 can be pointed at each in turn and the across-seed spread of the
FINAL metrics measured directly.