движе́ние
Classifies a Russian motion-verb candidate from its sentence context. The sense output distinguishes physical motion, non-motion use, another lemma, and unclear cases. A second output identifies hortative constructions such as Пойдём.
выраже́ния
Locates known multiword expressions with a pattern matcher, then classifies each occurrence as literal or idiomatic from its context. The model is not deployed in the browser.
Test split · 2,691 examples
| Class | Support | Precision | Recall | F1 |
|---|---|---|---|---|
| Physical motion | 1,773 | 94.36 | 96.33 | 95.34 |
| Non-motion use | 881 | 92.52 | 88.42 | 90.42 |
| Other lemma | 37 | 87.18 | 91.89 | 89.47 |
| Unclear | 0 | — | — | — |
Macro F1 includes classes with test support. The unclear class has no examples in this split.
TINY · BASE · LARGE
These models use the same v1 training data, task definition, and test split. The comparison isolates the effect of encoder size. The deployed model above uses newer data and distillation, so its score is reported separately.
| Size | Encoder | Layers | Width | Checkpoint | Sense accuracy | Sense macro F1 | Construction macro F1 |
|---|---|---|---|---|---|---|---|
| tiny | cointegrated/rubert-tiny2 | 3 | 312 | 116.8 MB | 83.88 | 68.83 | 72.29 |
| base | DeepPavlov/rubert-base-cased | 12 | 768 | 711.5 MB | 88.97 | 78.56 | 83.53 |
| large | ai-forever/ruRoberta-large | 24 | 1024 | 1422 MB | 94.63 | 86.75 | 91.56 |
v1 test split · 1,061 examples · values shown as percentages
NOT RELEASED
Experimental test result on 1,400 occurrences. The unseen subset contains 488 examples from expression categories excluded from training. Model selection and browser packaging are not final.