Text Classification
Transformers
Safetensors
Russian
bert
russian
sentiment-analysis
multi-class-classification
rubert
tiny
text-embeddings-inference
Instructions to use sergeyzh/rubert-tiny-sentiment with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use sergeyzh/rubert-tiny-sentiment with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-classification", model="sergeyzh/rubert-tiny-sentiment")# Load model directly from transformers import AutoTokenizer, AutoModelForSequenceClassification tokenizer = AutoTokenizer.from_pretrained("sergeyzh/rubert-tiny-sentiment") model = AutoModelForSequenceClassification.from_pretrained("sergeyzh/rubert-tiny-sentiment", device_map="auto") - Notebooks
- Google Colab
- Kaggle
File size: 4,781 Bytes
c8fa01e 8d2b14c c8fa01e 8d2b14c c8fa01e 8d2b14c | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 | ---
language:
- ru
pipeline_tag: text-classification
tags:
- russian
- sentiment-analysis
- multi-class-classification
- rubert
- tiny
- transformers
license: mit
base_model: sergeyzh/rubert-tiny-sts-v2
library_name: transformers
---
Компактная модель BERT-tiny для классификации сентимента русских отзывов: 3 класса — Negative (0), Neutral (1), Positive (2).
Получена на базе [sergeyzh/rubert-tiny-sts-v2](https://huggingface.co/sergeyzh/rubert-tiny-sts-v2) дистилляцией мягких меток (soft labels) от учителя [sergeyzh/rubert-large-uncased-sentiment](https://huggingface.co/sergeyzh/rubert-large-uncased-sentiment).
Основные характеристики модели:
- размер hidden — 312,
- длина контекста — 512,
- слоёв — 3,
- параметров — ~29M (вес ~111 МБ).
Классы: `0` — Negative, `1` — Neutral, `2` — Positive.
## Использование
Самый простой способ — `pipeline`:
```Python
from transformers import pipeline
model = pipeline("text-classification", model="sergeyzh/rubert-tiny-sentiment")
model("Просто шедевр. Каждая минута на вес золота, ни секунды скуки. Музыка, игра актёров, режиссура — всё на высочайшем уровне.", truncation=True, max_length=512)
# [{'label': 'Positive', 'score': 0.9313}]
```
Если нужны вероятности всех классов, используйте `transformers` напрямую:
```Python
import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification
checkpoint = "sergeyzh/rubert-tiny-sentiment"
tokenizer = AutoTokenizer.from_pretrained(checkpoint)
model = AutoModelForSequenceClassification.from_pretrained(checkpoint)
text = "Просто шедевр. Каждая минута на вес золота, ни секунды скуки. Музыка, игра актёров, режиссура — всё на высочайшем уровне."
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=512)
with torch.no_grad():
logits = model(**inputs).logits
proba = torch.softmax(logits, dim=1)
label = model.config.id2label[proba.argmax().item()]
print(label, proba.tolist())
# Positive [[0.0109, 0.0577, 0.9313]]
```
## Обучение
- Базовая модель: [sergeyzh/rubert-tiny-sts-v2](https://huggingface.co/sergeyzh/rubert-tiny-sts-v2)
- Метод: дистилляция мягких меток (soft-label distillation) от учителя [sergeyzh/rubert-large-uncased-sentiment](https://huggingface.co/sergeyzh/rubert-large-uncased-sentiment), α = 0.5
- Данные: 105500 русских отзывов (94950 train / 10550 val) — датасеты Kinopoisk, RuReviews, Georeview
- 3 эпохи, batch_size = 32 (grad_accum = 2), lr = 5e-5, warmup = 0.1, weight_decay = 0.01, max_length = 256
- Отбор по валидационной F1 (лучшая эпоха 2, val F1 = 0.7552)
## Метрики
Тестовые наборы: Kinopoisk (1500), RuReviews (15000), Georeview (5000, 5-звёздный рейтинг сведён к 3 классам: 1–2 звезды — Negative, 3 — Neutral, 4–5 — Positive). Оценка: argmax по логитам, max_length = 512.
| Модель | Kinopoisk Acc/F1 | RuReviews Acc/F1 | Georeview Acc/F1 | Avg F1 |
| :--- | :--- | :---: | :---: | :---: |
| [sergeyzh/rubert-large-uncased-sentiment](https://huggingface.co/sergeyzh/rubert-large-uncased-sentiment) | **0.7013** / **0.6929** | 0.7851 / 0.7866 | **0.7858** / **0.7361** | **0.7385** |
| **sergeyzh/rubert-tiny-sentiment** | 0.6593 / 0.6519 | 0.7672 / 0.7690 | 0.7680 / 0.7130 | 0.7113 |
| [seara/rubert-base-cased-russian-sentiment](https://huggingface.co/seara/rubert-base-cased-russian-sentiment) | 0.5653 / 0.5679 | **0.8163** / **0.8183** | 0.6566 / 0.6434 | 0.6765 |
| [seara/rubert-tiny2-russian-sentiment](https://huggingface.co/seara/rubert-tiny2-russian-sentiment) | 0.4980 / 0.5032 | 0.7877 / 0.7899 | 0.6218 / 0.6122 | 0.6351 |
| [blanchefort/rubert-base-cased-sentiment](https://huggingface.co/blanchefort/rubert-base-cased-sentiment) | 0.5253 / 0.5209 | 0.7615 / 0.7549 | 0.6716 / 0.6047 | 0.6268 |
| [blanchefort/rubert-base-cased-sentiment-rusentiment](https://huggingface.co/blanchefort/rubert-base-cased-sentiment-rusentiment) | 0.5413 / 0.5470 | 0.6230 / 0.6327 | 0.6022 / 0.5760 | 0.5852 |
| [cointegrated/rubert-tiny-sentiment-balanced](https://huggingface.co/cointegrated/rubert-tiny-sentiment-balanced) | 0.4293 / 0.3977 | 0.7330 / 0.7344 | 0.6158 / 0.5857 | 0.5726 |
|