ihsan31415/indo-roBERTa-financial-sentiment

🇮🇩 IndoRoBERTa for Indonesian Financial Sentiment Classification

This is a fine-tuned version of w11wo/indonesian-roberta-base-sentiment-classifier, specialized for Indonesian financial news sentiment classification since i cant find any financial sentiment models for indonesian market, i decided to make my self.

🧠 Model Summary

Field	Value
Model Name	ihsan31415/indo-roBERTa-financial-sentiment
Base Model	w11wo/indonesian-roberta-base-sentiment-classifier
Language	Indonesian (id)
Task	Sentiment Analysis (Financial)
Labels	0: Positive, 1: Neutral, 2: Negative (⚠️ flipped label order)
Dataset	intanm/indonesian-financial-sentiment-analysis + synthetic and augmented samples
Fine-tuned by	ihsan31415
Training Epochs	5 (Early stopping at epoch 5, best at epoch 3)
Eval Accuracy	97.49%

---

🧠 Model Objective

This model classifies Indonesian financial news articles into:

• 0 → Positive
• 1 → Neutral
• 2 → Negative

⚠️ Important: Label Mapping is Flipped This label order follows the base model's unexpected configuration. During training and evaluation, the dataset was relabeled accordingly.

⚠️ Always interpret model output using this mapping:

• 0: Positive
• 1: Neutral
• 2: Negative

---

📊 Dataset & Preprocessing Pipeline

🔹 Source Dataset

• intanm/indonesian-financial-sentiment-analysis
• Labeled financial news (imbalanced and limited)

📈 Data Augmentation & Balancing

1. 🧪 Gemini Synthetic Generation

• Generated structured financial news samples using gemini-2.0-flash-lite
• Targeted generation for underrepresented classes

2. ✍️ GPT-2 Prompt Completion

• Used indonesian-nlp/gpt2-medium-indonesian
• Prompt templates varied and strictly separated between train/test sets

3. 🧩 Roberta-Based Masked Augmentation

• Strategic masking/filling while protecting key financial terms
• Iterative masking to increase diversity and context coverage

📊 Final Label Distribution

Train Set:

2 (Negative): 22906
1 (Neutral): 23374
0 (Positive): 23423

Test Set:

2 (Negative): 9817
1 (Neutral): 10018
0 (Positive): 10039

---

🏋️ Training Details

🔁 Label Flipping

The base model uses non-standard labels:

• 0: Positive
• 1: Neutral
• 2: Negative

Training data was relabeled accordingly.

🔧 TrainingArguments

TrainingArguments(
    output_dir="./results-roberta",
    eval_strategy="epoch",
    save_strategy="epoch",
    logging_strategy="epoch",
    per_device_train_batch_size=256,
    per_device_eval_batch_size=256,
    num_train_epochs=15,
    learning_rate=2e-5,
    weight_decay=0.01,
    load_best_model_at_end=True,
    metric_for_best_model="accuracy",
    save_total_limit=4,
)

• Early stopping (patience=2)
• Training completed at epoch 5, best model from epoch 3

📊 Training Progress

Epoch	Training Loss	Validation Loss	Accuracy	Precision	Recall	F1 Score
1	0.104500	0.085562	0.969402	0.969715	0.969402	0.969356
2	0.029100	0.088392	0.974859	0.974914	0.974859	0.974860
3	0.012700	0.102305	0.974926	0.974949	0.974926	0.974933
4	0.008900	0.125707	0.972816	0.972959	0.972816	0.972846
5	0.004400	0.157659	0.966690	0.966902	0.966690	0.966676

✅ Evaluation Results

eval_loss                 = 0.10230540484189987
eval_accuracy             = 0.9749255130394028
eval_precision            = 0.9749490510899772
eval_recall               = 0.9749255130394028
eval_f1                   = 0.9749326327197978
eval_runtime              = 71.9098
eval_samples_per_second   = 415.395
eval_steps_per_second     = 1.627
epoch                     = 5.0

---

🔎 Usage

Using Pipeline

from transformers import pipeline

pretrained_name = "ihsan31415/indo-roBERTa-financial-sentiment"

nlp = pipeline(
    "sentiment-analysis",
    model=pretrained_name,
    tokenizer=pretrained_name
)

nlp("IHSG diprediksi melemah karena sentimen global negatif")

RAW

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

# Load model and tokenizer
model = AutoModelForSequenceClassification.from_pretrained("ihsan31415/indo-roBERTa-financial-sentiment")
tokenizer = AutoTokenizer.from_pretrained("ihsan31415/indo-roBERTa-financial-sentiment")

# Example input
text = "IHSG diprediksi melemah karena sentimen global negatif"
inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True)
outputs = model(**inputs)

# Get predicted class
predicted_label = torch.argmax(outputs.logits, dim=1).item()

# Interpret using flipped label mapping
label_map = {
    0: "Positive",
    1: "Neutral",
    2: "Negative"
}
print(f"Predicted sentiment: {label_map[predicted_label]}")

Author

this indonesian RoBERTa base financial CLassifier was trained and evaluated by Khoirul Ihsan using Google colab GPU T4.

---

📌 Citation

@misc{khoirul_ihsan_2025,
  title        = {IndoRoBERTa for Indonesian Financial Sentiment Classification},
  author       = {Khoirul Ihsan},
  howpublished = {\url{https://huggingface.co/ihsan31415/indo-roBERTa-financial-sentiment}},
  year         = {2025},
  note         = {Fine-tuned from w11wo/indonesian-roberta-base-sentiment-classifier using augmented financial news data from intanm/indonesian-financial-sentiment-analysis and various synthetic generation methods (Gemini, GPT-2, Roberta masking).},
  publisher    = {Hugging Face}
}

---

📬 Contact

Created with love and tears by ihsan:

For collaborations or questions, feel free to reach out via Hugging Face or GitHub.