Modelpublic

Bektur756/rubert-tiny2-russian-sentiment

sourceHugging Facemitupdated 4d agoView on Hugging Face

0likes21downloads

Model Card

Russian Sentiment Classifier

Классификатор тональности русскоязычных пользовательских текстов.

Classes

—0: Neutral
—1: Positive
—2: Negative

Base model

cointegrated/rubert-tiny2

Модель была полностью дообучена для классификации текста на три класса.

Dataset

MonoHime/ru_sentiment_dataset

После очистки и удаления дубликатов:

—Train: 180,853
—Validation: 9,642
—Test: 9,642

Training

—Epochs: 3
—Maximum sequence length: 128
—Learning rate: 2e-5
—Train batch size: 64
—Weighted Cross Entropy: enabled
—Best model metric: Negative F1
—GPU: Tesla T4

Class weights:

—Neutral: 1.3124
—Positive: 0.7020
—Negative: 1.2292

Test metrics

Metric	Value
Accuracy	0.7811
Macro Precision	0.7713
Macro Recall	0.7918
Macro F1	0.7768
Neutral F1	0.6960
Positive F1	0.7989
Negative F1	0.8354
Negative Recall	0.8787

Latency

End-to-end latency includes preprocessing, tokenization and inference. Batch size is 1.

Device	Mean	Median	P95	Requests/sec
Tesla T4 GPU	7.33 ms	6.57 ms	12.92 ms	136.41
Colab CPU	13.84 ms	13.69 ms	16.77 ms	72.26

Usage

python

import torch

from transformers import (
  AutoModelForSequenceClassification,
  AutoTokenizer,
)

MODEL_NAME = "Bektur756/rubert-tiny2-russian-sentiment"

tokenizer = AutoTokenizer.from_pretrained(
  MODEL_NAME
)

model = (
  AutoModelForSequenceClassification
  .from_pretrained(MODEL_NAME)
)

text = "Отличный сервис, всё понравилось."

inputs = tokenizer(
  text,
  return_tensors="pt",
  truncation=True,
  max_length=128,
)

with torch.inference_mode():
  logits = model(**inputs).logits

probabilities = torch.softmax(
  logits,
  dim=-1,
)[0]

predicted_id = int(
  probabilities.argmax()
)

print(model.config.id2label[predicted_id])
print(probabilities.tolist())
## Limitations

В исходном датасете присутствуют шумные и неоднозначные метки.
Слабее всего модель распознаёт нейтральный класс.