indobert-beauty-spam-classifier

A fine-tuned IndoBERT model for detecting spam in Indonesian-language beauty product reviews.

Model Details

• Base model: indolem/indobert-base-uncased
• Architecture: BertForSequenceClassification (12 layers, hidden size 768, 12 attention heads)
• Language: Indonesian (id)
• Task: Binary text classification (spam detection)
• Labels:
  • 0: Not Spam
  • 1: Spam

Intended Use

Classifies whether a beauty product review written in Indonesian is spam (promotional links, unrelated content, bot-like text) or a genuine review. Built for filtering user-generated review content on a beauty e-commerce platform.

Not intended for: languages other than Indonesian, review content outside the beauty/cosmetics domain, or as a sole moderation decision-maker in production without human review (see Limitations).

How to Use

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model_name = "sofyanwaldy/indobert-beauty-spam-classifier"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

text = "Dapatkan uang dengan cepat! Klik di sini untuk informasi lebih lanjut."
inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True)

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

predicted_class = torch.argmax(logits, dim=-1).item()
print("Spam" if predicted_class == 1 else "Not Spam")

Training Data

Fine-tuned on a custom dataset of 260 labeled Indonesian beauty product reviews (spam_classification_dataset_bahasa.csv):

• 162 not-spam examples
• 98 spam examples
• 80/20 train/test split (seed=42)

Training Procedure

Hyperparameter	Value
Epochs	3
Learning rate	2e-5
Batch size (train/eval)	8
Weight decay	0.01
Optimizer	AdamW (HF Trainer default)

Trained with the Hugging Face Trainer API on top of indolem/indobert-base-uncased.

Evaluation

No evaluation metrics were logged during training. Precision/recall/F1 on a held-out set have not yet been computed for this checkpoint — treat performance as unverified until evaluated.

Limitations and Bias

• Very small training set (260 examples): high risk of overfitting and poor generalization to review styles/topics not represented in the data.
• Class imbalance (~62% not-spam / 38% spam): may bias the model toward predicting "Not Spam."
• Domain-specific: trained only on beauty product reviews; unlikely to generalize to other review domains.
• No formal evaluation metrics available for this checkpoint — accuracy/F1 unknown.
• Not tested for adversarial or obfuscated spam patterns (e.g., leetspeak, character substitution).

License

MIT, inherited from base model indolem/indobert-base-uncased.