Skip to content

Commit a85a5eb

Browse files
author
Ronald Tse
committed
docs(verdicts): run-020 plane-distill 4.4462 - matched-supervision architecture win, label-limited data loss
1 parent f83bda2 commit a85a5eb

1 file changed

Lines changed: 14 additions & 0 deletions

File tree

‎docs/RESULTS.md‎

Lines changed: 14 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1025,3 +1025,17 @@ much as the knowledge. Paradigm-table aux is closed (no dose tuning,
10251025
no convention-normalization retry recorded as future work; a lexical
10261026
lever, if ever revisited, must be rendered into benchmark-convention
10271027
running text). The frontier mover on record remains the r7 news mix.
1028+
1029+
## run-020 plane-distill — label-limited distillation loses to raw-text scale (2026-10-05)
1030+
1031+
Arm B of the plane-large verdict follow-up: ByT5-small plane encoder
1032+
trained on r7 teacher labels (29,322 unique pairs ×6, 3 epochs) scores
1033+
full-set **Total DER 4.4462 / Morph 2.8620**. Two clean readings:
1034+
(a) on MATCHED teacher-label supervision the plane architecture beats
1035+
the seq2seq student (4.4462 vs 4.5701 — architecture advantage
1036+
confirmed independent of supervision); (b) label-limited distillation
1037+
loses to raw-text diversity at scale (run-018's 551k unique units →
1038+
3.5905) — for the plane family the lever is unique-data volume, not
1039+
supervision source. A full-corpus teacher labeling pass (~15-20h GPU)
1040+
is the conditional distill follow-up, only if the epoch-2 arm falls
1041+
short of the 2.2864 gate.

0 commit comments

Comments
 (0)