Study 260725091 — reported

30 done · partially reproduced

View study →
STUDY 260725091ARM 006 / 30
← Study record

Selected arm evidence

Pythia 160M × WikiText

Track R · seed 42 · R-pythia-160m-wikitext-s42

This arm tests the Pythia 160M and WikiText configuration under Track R at seed 42. The conditional release interval includes zero, so the directional result is inconclusive.

Arm directional labelINCONCLUSIVEOne arm; not the study classification.
Study classificationpartially reproducedFrozen across all 30 selected arms.
01

Published and rerun estimates

Interval inclusion and directional agreement are reported separately

Published reward Δ+0.0435Reported comparison point
Release-protocol reward Δ-0.1920Selected sampled endpoint
Conditional 95% interval[-0.6555, +0.2695]Fixed checkpoint and retained generations

Conditional interval comparison

Reward-difference estimates and conditional 95% intervals

Both estimates use the same horizontal scale; whisker length represents interval width.

The circle and square mark the release-sampled and paired-deterministic estimates. Horizontal lines show their supplied conditional 95% intervals. The dashed vertical line marks the published reward difference, and the solid vertical line marks zero. These intervals do not estimate training-to-training variance.
Numerical interval inclusionPublished estimate inside interval

This reports only whether the published point estimate falls inside the supplied conditional release interval.

Directional assessmentInconclusive because the conditional interval includes zero

The machine-readable arm record labels this result as inconclusive_interval_includes_zero; this arm-level assessment does not replace the study verdict.

Independent paired endpoint

Independent deterministic paired comparison

The Holm-adjusted p-value is calculated for this paired endpoint within its track. It does not apply to the sampled release endpoint.

Paired reward Δ
-0.1374
Conditional 95% interval
[-0.3689, +0.0872]
Within-track Holm p
1.0000