---
license: mit
language:
  - ru
base_model: cointegrated/rubert-tiny2
pipeline_tag: text-classification
tags:
  - rubert
  - bert
  - tiny
  - russian
  - classification
  - prompt-classification
  - intent-classification
---

# rubert-tiny2-clf

A fine-tuned version of [`RuBERT-tiny2`](https://huggingface.co/cointegrated/rubert-tiny2)
for **prompt intent classification** on Russian text: given a user prompt,
the model predicts which of three intents it corresponds to.

| Label | Meaning |
|---|---|
| `write` | The prompt asks to write/generate text |
| `draw` | The prompt asks to draw/generate an image |
| `neutral` | Neither of the above |

## Usage

```python
from transformers import pipeline

model = pipeline(model="r1char9/rubert-tiny2-clf")
model("Сгенерируй картину Томаса Шелби")
# [{'label': 'draw', 'score': 0.8699279427528381}]
```

## Metrics

Evaluated on a held-out test set (support: 291 examples total).

| Metric | write | draw | neutral | micro avg | macro avg | weighted avg |
|---|---|---|---|---|---|---|
| Precision | 1.0 | 1.0 | 1.0 | 1.0 | 1.0 | 1.0 |
| Recall | 1.0 | 1.0 | 1.0 | 1.0 | 1.0 | 1.0 |
| F1-score | 1.0 | 1.0 | 1.0 | 1.0 | 1.0 | 1.0 |
| Support | 155 | 117 | 19 | 291 | 291 | 291 |
| AUC-ROC | 1.0 | 1.0 | 1.0 | 1.0 | 1.0 | 1.0 |

> ⚠️ All metrics at 1.0 across every class and averaging scheme usually
> signals evaluation on data overlapping with training data, a very small /
> easy test set, or a label-leakage issue — worth double-checking with a
> truly held-out, more diverse test set before relying on this number.

## Limitations

- The `neutral` class has much lower support (19 examples) than `write`
  (155) and `draw` (117) — performance on the minority class may not be as
  reliable as the reported metrics suggest.
- The model is intended for short, single-intent prompts; behavior on long
  or multi-intent text is untested.