Skip to content
Merged
Show file tree
Hide file tree
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
2 changes: 1 addition & 1 deletion TODO.sota/00-README.md
Original file line number Diff line number Diff line change
Expand Up @@ -49,7 +49,7 @@ hf jobs run --flavor a100-large -d \
| 09 | 09-hebrew-audio-pseudolabel-probe.md | P3 (parked probe) | — | DONE as parked probe (script only) |
| 10 | 10-heb-plane-2-release.md | P1 | 04 | DONE (release, index-v9, 3 pin bumps, card, golden, HF mirror #28) |
| 11 | 11-arabic-r8-chain.md | P1 | — | jobs RUNNING (a100) |
| 12 | 12-thai-tiny-contract.md | P2 | — | job RUNNING (a100) |
| 12 | 12-thai-tiny-contract.md | P2 | — | NEGATIVE (caveated) — PER 2038%; sub-10MB closed |
| 13 | 13-thai-lexicon-release.md | P2 | 03 | DONE (tha-lexicon-kaikki-1.0 + sha, CC BY-SA) |
| 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) |
Expand Down
15 changes: 15 additions & 0 deletions docs/RESULTS.md
Original file line number Diff line number Diff line change
Expand Up @@ -1270,3 +1270,18 @@ 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.

## WO12 Thai tiny tier — NEGATIVE (caveated): the 12M student misses by an order of magnitude (2026-10-08)

run-025 (three attempts: two decode harness bugs — device mix, vocab
scope — fixed train#113/#114; third attempt trained cleanly to loss
~0.30 then scored **PER 2038%** greedy on the kaikki gate, followed by
a third device bug in the export step). Verdict: the gate (≤3.5 PER,
<10MB) FAILED as measured. Two honest readings, both recorded:
(a) the in-job greedy decode was never unit-tested and shows runaway
repetition — a decode rewrite could change the number; (b) even a
perfect decode cannot rescue a 12M-param student to 2.85%-class Thai
phonotactics from ~200K teacher pairs. Sub-10MB tier CLOSED unless the
owner wants a tested-decode rerun. The shipped Thai tier remains
tha-g2p-small (219MB, 2.92% model-only) + the hybrid (0.1478 PER
@0.73ms) — the client story is already won at the dictionary tier.
Loading