---
language: ru
license: mit
library_name: transformers
pipeline_tag: text-classification
base_model: DeepPavlov/rubert-base-cased
tags:
  - intent-classification
  - russian
  - aitu
metrics:
  - f1
  - accuracy
---

# AITU Student Assistant — Intent Classifier (rubert-base-cased)

Russian intent classifier for an AITU student-assistant chatbot. Maps a student
message to one of **45 intents** across 8 domains (schedule, assignments,
grades, exams, documents, payments, campus, general).

## Results (held-out test, 1,790 messages)

| metric | value |
|---|---|
| macro-F1 | **0.964** |
| accuracy | 0.963 |
| weighted-F1 | 0.963 |

Fine-tuned from `DeepPavlov/rubert-base-cased` on 17,898 cleaned messages
(stratified 80/10/10, class-weighted cross-entropy, 6 epochs).

## Usage

```python
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

m = "govnejri/aitu-intent-rubert-base-cased"
tok = AutoTokenizer.from_pretrained(m)
model = AutoModelForSequenceClassification.from_pretrained(m)

text = "когда дедлайн оплаты за обучение"
enc = tok(text, return_tensors="pt", truncation=True, max_length=96)
probs = torch.softmax(model(**enc).logits, -1)[0]
top = int(probs.argmax())
print(model.config.id2label[top], float(probs[top]))  # payment_deadline ~0.95
```

Educational Practice 2025-2026 project. Trained on synthetic data — validate on
real traffic before production.