Skip to content

docs: verdict wave 2 — dose-response mapped #276

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
ronaldtse merged 1 commit into from
Oct 7, 2026
Merged

docs: verdict wave 2 — dose-response mapped #276

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
2 changes: 1 addition & 1 deletion TODO.sota/00-README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -54,7 +54,7 @@ hf jobs run --flavor a100-large -d \
| 14 | 14-ts-plane-port.md | P2 | — | DONE (ts#99/#100/#101, CI leg models#265) |
| 15 | 15-api-edge.md | P2 | 14 | DONE (api#29: edge-first /v1/infer, kind in index) |
| 16 | 16-parking-lot.md | — | — | parked (not now) |
| 17 | 17-arabic-register-closure.md | P1 | 11 | r8a RUNNING; r8b launching |
| 17 | 17-arabic-register-closure.md | P1 | 11 | dose-response MAPPED (RESULTS wave 2); r8d knee-seeking RUNNING; specialist option staged |
| 18 | 18-parity-runs.md | P1 | 10 | pending |
| 19 | 19-lexicon-loader.md | P2 | 13 | DONE (py#33) |
| 20 | 20-hebrew-large-arm.md | P1 | 04 | NEGATIVE — 8.48 vs base 8.18; base stays shipped |
Expand Down
23 changes: 23 additions & 0 deletions docs/RESULTS.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1247,3 +1247,26 @@ variant (r8d) can hold ID while keeping the OOD gain.
**8.48** vs base's 8.31 (same protocol). 1.2B params on 50K units
saturates; base stays shipped at 8.18. Capacity is not the Hebrew
constraint at this data scale — data is.

## WO17 verdict wave 2 — the dose-response curve mapped; r8d launched at the knee (2026-10-08)

r8b (QCRI silver, 33% dose) completes the curve. All arms r7-init,
1 epoch, WINDOW=600, dual-surface gate:

| arm | silver dose | SadeedDiac-25 DER (ID) | WikiNews-2024 multiref WER/DER (OOD) |
|---|---|---|---|
| r7 baseline | 0 | **2.2864** | 17.38 / 11.83 |
| r8a (arwiki self) | 2.7% | 2.9087 | 16.71 / 10.83 |
| **r8b (QCRI)** | 33% | 4.3911 | **11.03 / 9.24** |

Readings: (1) news-register silver moves OOD monotonically and STEEPLY —
r8b takes **6.35 WER / 2.6 DER** off the OOD number, the largest
single-lever move of the campaign; (2) the ID cost is also monotone —
the frontier is Pareto, not a winner; (3) r8b doubles as a strong
NEWS-DOMAIN SPECIALIST (11.03 OOD, approaching QCRI's in-domain 2.70
where our 17.38 was measured).

Actions: **r8d launched** at the presumed knee (arwiki, 5K highest-keep
windows, lr 2e-5 — dose knobs train#115) targeting OOD ~13-15 with ID
≤2.5. Register-specialist ship option staged (WO24): ara-diac-news-1.0
alongside ara-diac-2.0, two-families doctrine. r8c (both) pending.
Loading