README.md

---
license: mit
language:
- id
metrics:
- accuracy
- f1
base_model: indobenchmark/indobert-base-p1
pipeline_tag: text-classification
library_name: transformers
tags:
- indoBERT
- classification
- aduan
- indonesian
model-index:
- name: aduan-model
  results:
  - task:
      type: text-classification
      name: Text Classification
    dataset:
      name: Custom Labeled Aduan Dataset
      type: private
      split: validation
    metrics:
    - type: accuracy
      value: 0.9389
    - type: f1
      value: 0.9389
---

# 📊 Indonesian Complaint Classification Model (IndoBERT)

[![Model](https://img.shields.io/badge/🤗%20Hugging%20Face-Model-yellow)](https://huggingface.co/Zulkifli1409/aduan-model)
[![License](https://img.shields.io/badge/License-Apache%202.0-blue.svg)](https://opensource.org/licenses/Apache-2.0)
[![Language](https://img.shields.io/badge/Language-Indonesian-red.svg)](https://en.wikipedia.org/wiki/Indonesian_language)

Model klasifikasi teks aduan masyarakat dalam Bahasa Indonesia menggunakan **IndoBERT (indobenchmark/indobert-base-p1)**.  
Model dapat mengelompokkan aduan ke dalam **5 kategori** dengan akurasi **96.10%**.

---

## 📑 Kategori Klasifikasi

| Label | Deskripsi | Contoh |
|-------|-----------|--------|
| **PINALTI** | Konten yang mengandung kata kasar, SARA, pornografi, ujaran kebencian, atau pelanggaran norma | "Kampret pejabat koruptor!", "Konten porno beredar", "Rasis banget pemerintah" |
| **DARURAT** | Situasi darurat yang membutuhkan respon segera (kebakaran, kecelakaan, bencana, ancaman nyawa) | "Ada kebakaran besar di pasar!", "Kecelakaan beruntun di tol", "Banjir bandang melanda desa" |
| **PRIORITAS** | Permasalahan yang perlu penanganan cepat (infrastruktur rusak, kebersihan, pelayanan publik) | "Jalan berlubang berbahaya", "Sampah menumpuk seminggu", "Lampu jalan mati semua" |
| **UMUM** | Pertanyaan informasi, saran, atau aduan non-urgent | "Bagaimana cara mengurus KTP?", "Kapan jadwal posyandu?", "Saran untuk program desa" |
| **LAINNYA** | Aduan yang tidak termasuk kategori di atas | "Terima kasih atas pelayanannya", "Hanya ingin menyampaikan apresiasi" |

---

## 🎯 Model Performance

### **Overall Metrics**
- **Validation Accuracy**: **96.10%**
- **Macro F1-Score**: **0.9608**
- **Weighted F1-Score**: **0.9610**
- **Average Confidence**: **93.90%**

### **Per-Class Performance**

| Label | Precision | Recall | F1-Score | Support |
|-------|-----------|--------|----------|---------|
| Pinalti | 0.9588 | 0.9645 | 0.9617 | 169 |
| Darurat | 0.9453 | 0.9603 | 0.9528 | 126 |
| Prioritas | 0.9675 | 0.9675 | 0.9675 | 123 |
| Umum | 0.9752 | 0.9593 | 0.9672 | 123 |
| Lainnya | 0.9596 | 0.9500 | 0.9548 | 100 |

### **Confusion Matrix**
```
                Predicted
              Pin  Dar  Pri  Umu  Lai
Actual  Pin   163   2   1   0   3
        Dar     2 121   2   0   1
        Pri     0   3 119   1   0
        Umu     2   2   1 118   0
        Lai     3   0   0   2  95
```

---

## 📊 Dataset Information

- **Total Training Samples**: 3,204
  - Pinalti: 844
  - Darurat: 630
  - Prioritas: 612
  - Umum: 616
  - Lainnya: 502
- **Train/Val Split**: 80% / 20% (2,563 / 641)
- **Augmentation**: Applied to balance classes
- **Language**: Indonesian (Bahasa Indonesia)

---

## 🚀 Quick Start

### Installation
```bash
pip install transformers torch
```

### Basic Usage
```python
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

# Load model and tokenizer
model_name = "Zulkifli1409/aduan-model"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

# Prepare input
text = "Ada kebakaran besar di pasar, tolong kirim pemadam segera!"
inputs = tokenizer(text, return_tensors="pt", padding=True, truncation=True, max_length=128)

# Predict
with torch.no_grad():
    outputs = model(**inputs)
    probs = torch.nn.functional.softmax(outputs.logits, dim=1)
    pred_idx = torch.argmax(probs).item()

# Labels
labels = ["PINALTI", "DARURAT", "PRIORITAS", "UMUM", "LAINNYA"]

print(f"Prediksi: {labels[pred_idx]}")
print(f"Confidence: {probs[0][pred_idx].item():.2%}")
print(f"\nAll probabilities:")
for label, prob in zip(labels, probs[0]):
    print(f"  {label}: {prob.item():.2%}")
```

**Output:**
```
Prediksi: DARURAT
Confidence: 96.03%

All probabilities:
  PINALTI: 0.21%
  DARURAT: 96.03%
  PRIORITAS: 2.89%
  UMUM: 0.45%
  LAINNYA: 0.42%
```

---

## 🧪 Example Predictions

| Input Text | Prediction | Confidence |
|------------|------------|------------|
| "Brengsek! Pejabat korup semua!" | **PINALTI** | 94.23% |
| "Ada orang kecelakaan parah butuh ambulans" | **DARURAT** | 95.67% |
| "Jalan berlubang perlu diperbaiki segera" | **PRIORITAS** | 92.34% |
| "Bagaimana cara mengurus surat izin usaha?" | **UMUM** | 89.45% |
| "Terima kasih atas bantuannya" | **LAINNYA** | 88.91% |
| "Konten porno tersebar di grup WhatsApp" | **PINALTI** | 91.78% |
| "Banjir tinggi merendam rumah warga" | **DARURAT** | 93.12% |
| "Sampah menumpuk di jalan sejak seminggu lalu" | **PRIORITAS** | 90.56% |

---

## 🔧 Batch Prediction

```python
texts = [
    "Ada kebakaran di gedung!",
    "Jalan rusak parah",
    "Dasar bodoh pemerintah!",
    "Kapan jadwal vaksinasi?"
]

# Tokenize batch
inputs = tokenizer(texts, return_tensors="pt", padding=True, truncation=True, max_length=128)

# Predict
with torch.no_grad():
    outputs = model(**inputs)
    probs = torch.nn.functional.softmax(outputs.logits, dim=1)
    predictions = torch.argmax(probs, dim=1)

labels = ["PINALTI", "DARURAT", "PRIORITAS", "UMUM", "LAINNYA"]

for text, pred_idx, prob in zip(texts, predictions, probs):
    pred_label = labels[pred_idx]
    confidence = prob[pred_idx].item()
    print(f"Text: {text}")
    print(f"Prediction: {pred_label} ({confidence:.2%})\n")
```

---

## 🌐 API Deployment

Model ini juga tersedia sebagai REST API di Railway:

**Base URL**: `https://api-klasifikasi-aduan.up.railway.app`

### cURL Example
```bash
curl -X POST https://api-klasifikasi-aduan.up.railway.app/predict \
  -H "Content-Type: application/json" \
  -d '{"text": "Ada kebakaran di pasar"}'
```

### Response
```json
{
  "label": "DARURAT",
  "confidence": 0.9603,
  "all_scores": {
    "PINALTI": 0.0021,
    "DARURAT": 0.9603,
    "PRIORITAS": 0.0289,
    "UMUM": 0.0045,
    "LAINNYA": 0.0042
  }
}
```

---

## 🛠️ Training Details

### Model Architecture
- **Base Model**: `indobenchmark/indobert-base-p1`
- **Task**: Sequence Classification (5 classes)
- **Max Sequence Length**: 128 tokens
- **Hidden Size**: 768
- **Attention Heads**: 12
- **Layers**: 12

### Training Configuration
- **GPU**: Tesla T4 (14.74 GB VRAM)
- **Precision**: FP16 (Mixed Precision)
- **Gradient Checkpointing**: Enabled
- **Batch Size**: 2
- **Learning Rate**: 1.5e-5
- **Epochs**: 5
- **Optimizer**: AdamW
- **Best Epoch**: 5

### Training Progress
| Epoch | Train Loss | Train Acc | Val Loss | Val Acc | Val F1 |
|-------|------------|-----------|----------|---------|--------|
| 1 | 0.3688 | 74.87% | 0.0825 | 93.45% | 0.9346 |
| 2 | 0.0586 | 95.86% | 0.0604 | 96.10% | 0.9609 |
| 3 | 0.0179 | 98.52% | 0.0635 | 96.41% | 0.9641 |
| 4 | 0.0069 | 99.38% | 0.0668 | 96.10% | 0.9611 |
| 5 | 0.0021 | 99.88% | 0.0623 | **96.10%** | **0.9610** |

---

## ⚠️ Important Notes

### Content Moderation (PINALTI)
Model ini dapat mendeteksi konten yang tidak pantas, namun **tidak sempurna**. Untuk aplikasi produksi yang sensitif, pertimbangkan:
- Layer moderasi tambahan
- Human review untuk kasus borderline
- Whitelist/blacklist kata kunci eksplisit
- Kombinasi dengan rule-based filtering

### Limitations
- Model dilatih dengan data aduan masyarakat Indonesia
- Performa optimal untuk teks dengan panjang 10-100 kata
- Slang atau dialek daerah tertentu mungkin kurang akurat
- Context yang ambigu dapat menghasilkan prediksi yang kurang tepat

---

## 📄 License

This model is licensed under **Apache 2.0 License**.

---

## 📧 Citation & Contact

**Developer**: Zulkifli1409  
**Hugging Face**: [@Zulkifli1409](https://huggingface.co/Zulkifli1409)

Jika Anda menggunakan model ini dalam penelitian atau aplikasi, mohon untuk memberikan kredit yang sesuai.

### BibTeX
```bibtex
@misc{zulkifli2025aduan,
  author = {Zulkifli},
  title = {Indonesian Complaint Classification Model with IndoBERT},
  year = {2025},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/Zulkifli1409/aduan-model}}
}
```

---

## 🤝 Contributing

Umpan balik, laporan bug, dan kontribusi sangat diterima!  
Silakan buka *issue* di repository atau hubungi via Hugging Face.

---

**© 2025 - Klasifikasi Aduan Model**