mSatashi/mentilinSafe-IndoDiscourse-harm-Multiclass

mentilinSafe-BERT-Multiclass

Model ini adalah fine-tuned version dari Exqrch/IndoDiscourse-ToxicityClassifier (BERT) yang dikonfigurasi ulang untuk Multi-class Classification guna mendeteksi 19 jenis kategori bahaya (harm) dalam bahasa Indonesia berdasarkan dataset IndoSafety.

Detail Model

• Base Model: BERT (IndoDiscourse-ToxicityClassifier)
• Tugas: Multi-class Text Classification
• Jumlah Kategori: 19 Kategori
• Akurasi Evaluasi: 95.39%

Kategori yang Didukung

Model ini dapat mengklasifikasikan teks ke dalam kategori berikut:

• Adult Content
• Assisting illegal activities
• Causing material harm by disseminating misinformation e.g. in medicine or law
• Compromise privacy by leaking or inferring private information (person/individual)
• Disseminating false or misleading information
• Ethnicities and Cultural Practices
• Historical Controversies
• Indonesian Entities
• Mental Health or Overreliance Crisis
• Nudging or advising users to perform unethical or unsafe actions
• Pancasila Misinterpretation and Corruption
• Reducing the cost of disinformation campaigns
• Regional Separatism Advocacy
• Religions and Beliefs
• Risks from leaking or inferring sensitive information (organization/gov)
• Social stereotypes and unfair discrimination
• Supernatural
• Toxic language (hate speech)
• Treat Chatbot as a Human

Cara Penggunaan

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model_name = "mSatashi/mentilinSafe-BERT-Multiclass"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

def predict(text):
    inputs = tokenizer(text, return_tensors="pt", padding=True, truncation=True)
    with torch.no_grad():
        outputs = model(**inputs)
        probs = torch.nn.functional.softmax(outputs.logits, dim=-1)
        pred_id = torch.argmax(probs, dim=-1).item()
    return model.config.id2label[pred_id], probs[0][pred_id].item()

text = "Masukkan kalimat di sini"
label, score = predict(text)
print(f"Kategori: {label} ({score:.4f})")

Hasil Evaluasi

Berdasarkan pengujian pada dataset IndoSafety-Eval-1:

• Accuracy: 0.9539
• Loss: 0.1924