atasoglu/turkish-base-bert-uncased-offenseval2020_tr

This is an offensive language detection model fine-tuned with coltekin/offenseval2020_tr dataset on ytu-ce-cosmos/turkish-base-bert-uncased.

Usage

Quick usage:

from transformers import pipeline
pipe = pipeline("text-classification", "atasoglu/turkish-base-bert-uncased-offenseval2020_tr")
print(pipe("bu bir test metnidir.", top_k=None))
# [{'label': 'NOT', 'score': 0.9970345497131348}, {'label': 'OFF', 'score': 0.0029654440004378557}]

Or:

import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model_id = "atasoglu/turkish-base-bert-uncased-offenseval2020_tr"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id).to(device)

@torch.no_grad
def predict(X):
    inputs = tokenizer(X, padding="max_length", truncation=True, max_length=256, return_tensors="pt")
    outputs = model.forward(**inputs.to(device))
    return torch.argmax(outputs.logits, dim=-1).tolist()

print(predict(["bu bir test metnidir."]))
# [0]

Test Results

Test results examined on the test split of fine-tuning dataset.

	precision	recall	f1-score	support
NOT	0.9162	0.9559	0.9356	2812
OFF	0.7912	0.6564	0.7176	716

accuracy			0.8951	3528
macro avg	0.8537	0.8062	0.8266	3528
weighted avg	0.8908	0.8951	0.8914	3528