Skip to content
Merged
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
19 changes: 19 additions & 0 deletions TODO.sota/39-ar-oracle-distill.md
Original file line number Diff line number Diff line change
Expand Up @@ -77,3 +77,22 @@ timeout — resume6/8 died exactly there). Stats:
sft stage planned it (O_* envs, both mounts) and the Runner's budget
guard cleared it — the first dogfooded launch; hand-typed job commands
retire here. Gate: first-alt WER < 3.5607; stretch ≤ 2.70.

## FINAL (2026-10-19): stage-2 REJECTED — ara canonical unchanged

oepoch-1 dual-surface eval (the training itself converged, loss 0.002;
eval initially crashed on missing pandas — rerun eval-only):

- SadeedDiac-25: Total DER **8.32** (gate: hold r7's 2.29) — identity
COLLAPSE
- WikiNews-2024 multiref: RAW 9.95 WER / 9.06 DER; first-alt
**3.97 WER** (gate < 3.5607) — MISS

The stage-1 oracle was real (0.54% relabel CER) but continue-training
on the relabeled 290K subset regressed both surfaces — domain-narrowed
distillation destroyed general calibration. Same family of negative as
fas stage-2. **ara canonical: ara-diac-2.0 (shipped, DER 2.29); the
oracle-distillation lever is now data-closed for BOTH ara and fas.**

Ops scars recorded: spine deps gap (fixed, #303), eval-only rerun
pattern (eval_oepoch1.py), eval extras belong in recipe deps.
Loading