мм01—03

техническийотдел

runlp

Small models.
The thinking
behind them.

A technical companion to the language lab. What each model does, how it works, and how we measure it.

01 / МОДЕЛИ

A question of language.

Three focused tasks.
One small language lab.

движе́ние

VERBS OF MOTION

Is something actually moving? A contextual classifier separates physical motion from figurative uses and words with the same spelling.

хожу в парк → physical motion

MODEL DETAILS

выраже́ния

MULTIWORD EXPRESSIONS

Which words belong together? Multiword expressions, phrasal units, and idioms often carry a meaning that cannot be read one word at a time.

бить баклуши → to loaf around

COMING VERY SOON

ударе́ние

WORD STRESS

Which vowel carries the stress? Marking it makes pronunciation visible and helps distinguish words that share a spelling.

язык → язы́к

PREPARED EXAMPLE

The motion classifier runs locally in the main-page playground. Multiword expressions are coming next, followed by stress.

02 / АРХИТЕКТУРА · ММ–01

Context does
the work.

The same verb.
A different meaning.

Он идёт домой.PHYSICAL MOTION / HE IS GOING HOME

Дождь идёт.NON-MOTION USE / IT IS RAINING

A dictionary finds the candidate. A small encoder reads the surrounding sentence. Two classification heads determine the sense and whether the verb invites someone to act.

Dictionary rules supply the verb’s structure; the model supplies the contextual judgement.

Text → meaningFIG. 01 / SIGNAL PATH

v3p-tiny-distilledONNX / FP32

01 / HIDDEN WIDTH312features

02 / CONTEXT WINDOW192tokens

03 / TEST ACCURACY93.68%

INSPECT RESULTS

THREE STAGES / ONE PREDICTIONSELECT A STAGE TO INSPECT

One small encoder. A contextual representation.

RuBERT-tiny2 encodes up to 192 tokens. The [CLS] state and the target-marker state are joined into a 624-wide representation for the two classification heads.

312 + 312 → 624

03 / ИЗМЕРЕНИЯ

Measured performance.

EXPORT REPORT

SENSE CLASSIFICATION / F1

0F1 SCORE (%)100

motionn = 1,773

95.34

non-motionn = 881

90.42

other-lemman = 37

89.47

unclearn = 0

N/A

“Unclear” has no examples in this split. Macro F1 includes supported classes only.

ACCURACY93.68%

MACRO F191.75%

SAMPLES2,691

TEST CONDITIONS & METHODOLOGY +

These are saved offline evaluation results for the v3p dataset, not a measurement of this browser. The student uses RuBERT-tiny2 with distillation from the v3-ruroberta-large teacher. The checkpoint was selected by development-set sense macro F1.

Test and development splits have different class balances. The construction test contains 2,658 “none” and 33 “hortative” examples; its high accuracy should be read alongside macro F1 (95.00%). Browser speed depends on your hardware and execution backend.

Source: model export manifest and training metrics · release d23ce01cc95c2fd9f9c8

FROM THEORY TO PRACTICETry the models.

The interactive language lab lives on the main page.