MovaRelease 0.4 — notes (published 2026-10-16)
0.4 is the release of the school: one parser model trained with the school's own agreed trees, a maths student that reads word problems through the dependency tree, primary-school courses, and a GPU mode for the heavy batch stages. The short list of changes is in Changelog.
What is new
- One model,
en.bin. Theentagger and parser are trained on UD English EWT, ESL Spoken and the agreed school treebank (42 759 sentences that several parsers and a teacher check agreed on). The model is not in git:install.shdownloads it from https://research.elamur.ai/download/0/0.4/ and checks its SHA-256. It is under CC BY-SA 4.0, as the treebanks it learns from. - The tree reader for word problems (
world/src/school_tree.rs,world treeread). Told numbers, actors, verbs, receivers and comparisons are read off the UD tree into the maths student's world. It replaced the old reader of word windows. On the course lessons (2026-10-15): lessons 1–38 and 49–64 2367 → 2380, the gate 2443 → 2733, lessons read as data 1123 → 1338, the exam 627 → 635; the two-fold checkcvon SVAMP train, world reader 297 → 513. - Primary-school courses (school.md): 25 knowledge courses with lessons, exercises, tests and exams, 26 992 test questions, and the pupils' dictionary. The level is a clever child of 9–10; half of every conversation is plain English and half is slang and regional speech.
- Two execution modes (gpu-exec.md). CPU only is the default and is exact. GPU
enhanced (
--features gpu,MOVA_EXEC=gpu) runs the batch stages on the GPU; each stage is either exact (gated byte for byte against the CPU) or an approximant whose agreement is measured. induceandselfplayon the GPU (training.md): break counting is an exact integer kernel, so the rules and reports are the same as on the CPU.- One Cargo workspace.
cargo build --releaseat the root builds everything intotarget/release/. - Corpus 0.4 (https://mova.elamur.ai/free/en/releases/0.4/): the school's books, the agreed school treebank, silver and bronze trees, the pupils' dictionary, maths scripts and the contrast sets, all under CC BY-SA 4.0.
- Licences. The code is under the Elamur Cosmic License (Elamur Cosmic License, licence/); the model and the corpus are under CC BY-SA 4.0. To be reviewed in 0.5.
Metrics 0.3 → 0.4
The same commands as in Tests, run on a CPU with the workspace build and
en.bin (sha256 31a93401fe7135b17e1eacccf608d1a42e94ebc8951f011d7f80ff9db144eb86). Held-out tests were run once.
| measure | 0.3 (2026-10-03) | 0.4 (2026-10-16) |
|---|---|---|
| bAbI test, 20 tasks | 99.8%, 20 of 20 ≥ 95% | 99.4%, 19 of 20 (task 18 92.3%, task 16 95.3%) |
| bAbI contrast set | 99.7%, 20 of 20 | 99.4%, 19 of 20 |
| StepGame, original labels | 96.3% | 96.3% |
| StepGame corrected, clean and noise | 100% | 100% |
| StepGame contrast set (no lexical reader) | 100% (17.1%) | 100% (17.1%) |
| SpartQA-Human YN / FR / total | 75.5 / 63.6 / 74.1 | 75.5 / 63.6 / 74.1 |
SVAMP test, math solver | 64.0% (192) | 64.3% (193) |
| SVAMP test, the school's maths student | 169 (2026-10-08, between releases) | 160 |
maths student cv on SVAMP train: world reader / book's way | 297 / 1116 (before the tree reader) | 513 / 1116 |
| LAS, tales test (with the shipped repair rules) | 77.38 | 79.69 |
| LAS, UD English EWT test | 79.86 | 80.68 |
| primary-school courses, 26 992 test questions | — | 11 730 right (43.5%) |
Notes:
- bAbI fell because of the parser model, not the code. With the 0.3 model
ud-ewt-eslspok.binthe same 0.4 code gives 99.8% and 20 of 20 (task 18 96.4%, task 16 99.5%).en.bintakes Does for the main verb in a question with a long subject ("Does the box of chocolates fit in the box?"). This is the main known issue of 0.4. - The school's SVAMP test fell by 9 (Subtraction 95 → 87, Common-Division 39 → 36, Addition 24 → 26,
Multiplication 11 → 11), while
cvon train did not fall. The probable cause is the parser change; not checked. - LAS rose on both tests. No sentence of the tales test or of the EWT test is in the model's training data (checked by text against the agreed school treebank); the school text is close in genre to the tales.
- Determinism. The bAbI test was run twice; the outputs are identical.
GPU mode (2026-10-15)
| stage | CPU | GPU | agreement |
|---|---|---|---|
parsing the 25 courses (gparse instead of en) | 11 730 right | 11 817 right | an approximant: other trees, deterministic run to run |
pragmatic witnesses (WITNESS_FAST=1) | 3 min 16 s | 59 s | labels agree on 99.5–100%; 3 of 2 200 differ after agreement |
induce: counting broken sentences, 152 000 sentences | 1× | 34× faster | exact: the same rules byte for byte |
selfplay round | 16.7 s | 6.9 s | exact |
Known issues
- Do-support questions with a long subject (above).
- The maths reader has no verb class for give away:
en.binnow reads it as a phrasal verb (compound:prt, which is right), and "she gives away 64 games" is not read. The unit testquestion_numbers_must_be_usedis marked ignored until 0.5. - The school's maths student is far below the
mathsolver on the SVAMP test (160 against 193): it reads through a world of actors, which is slower to learn but explains every step.
Not carried on
- The 0.3 model
ud-ewt-eslspok.bin(it stays in the 0.3 download). - The old maths reader of word windows.
- The story worlds without a teacher's check, and the unfinished courses (programming, philosophy).
This page in Markdown → https://md.mova.elamur.ai/free/en/docs/releases/0.4/