README.md

metadata

license: mit
language:
  - id
library_name: transformers
tags:
  - sentiment-analysis
  - finance
  - indonesian
  - roberta
  - text-classification
datasets:
  - intanm/indonesian-financial-sentiment-analysis
  - indonlp/indonlu
  - triagungj/cnbc-indonesia-stock-news-sentiment-dataset
metrics:
  - accuracy
  - f1
  - precision
  - recall
base_model: ihsan31415/indo-roBERTa-financial-sentiment
pipeline_tag: text-classification
model-index:
  - name: indo-roBERTa-financial-sentiment-v2
    results:
      - task:
          type: text-classification
          name: Sentiment Analysis
        metrics:
          - type: accuracy
            value: 0.9284
          - type: f1
            value: 0.9284
          - type: precision
            value: 0.9284
          - type: recall
            value: 0.9284

IndoRoBERTa Financial Sentiment v2

Fine-tuned Indonesian financial sentiment classifier built on top of ihsan31415/indo-roBERTa-financial-sentiment, which itself derives from w11wo/indonesian-roberta-base-sentiment-classifier.

Model Details

Field	Value
Base Model	ihsan31415/indo-roBERTa-financial-sentiment
Architecture	RoBERTa (~125M params)
Language	Indonesian (Bahasa Indonesia)
Task	3-class Financial Sentiment Classification
License	MIT
Fine-tuned on	Google Colab T4 GPU

Label Mapping

⚠️ Non-standard label order — inherited from the base model.

Label ID	Sentiment
0	Positive
1	Neutral
2	Negative

Training Data

This model was fine-tuned on a combination of three datasets, augmented with synthetic data:

Source	Type	Samples
intanm/indonesian-financial-sentiment-analysis	HuggingFace Dataset	Primary source
CNBC Indonesia Stock News Sentiment	Kaggle (2024-2025)	9,819 headlines
IndoNLU SmSA	HuggingFace (IndoNLU)	~12,000 reviews

Data Augmentation

Three augmentation methods were used to balance class distribution:

1. Gemini Synthetic Generation (gemini-2.5-flash) — Generated structured financial news sentences targeting underrepresented classes
2. GPT-2 Prompt Completion (indonesian-nlp/gpt2-medium-indonesian) — Diverse paraphrasing from financial prompt templates
3. RoBERTa Masked Augmentation — Strategic token masking with protected financial keywords (tickers, directional terms)

Final Dataset Distribution

Split	Samples
Train	20,616
Test	5,155

Training Configuration

TrainingArguments(
    per_device_train_batch_size=64,
    gradient_accumulation_steps=4,       # Effective batch size: 256
    num_train_epochs=15,                 # With early stopping (patience=3)
    learning_rate=1e-5,                  # Lower LR for continued fine-tuning
    weight_decay=0.01,
    warmup_ratio=0.1,
    lr_scheduler_type="cosine",
    metric_for_best_model="f1",
    fp16=True,
)

Training Progress

Epoch	Train Loss	Val Loss	Accuracy	F1	Precision	Recall
1.0	0.8945	0.4876	0.7996	0.7998	0.8001	0.7996
2.0	0.4094	0.3435	0.8640	0.8634	0.8658	0.8640
3.0	0.2982	0.2725	0.8980	0.8979	0.8981	0.8980
4.0	0.2253	0.2469	0.9115	0.9113	0.9117	0.9115
5.0	0.1785	0.2418	0.9187	0.9186	0.9188	0.9187
6.0	0.1475	0.2384	0.9232	0.9231	0.9233	0.9232
7.0	0.1275	0.2450	0.9214	0.9213	0.9215	0.9214
8.0	0.1069	0.2452	0.9251	0.9251	0.9251	0.9251
9.0	0.0904	0.2465	0.9259	0.9258	0.9258	0.9259
10.0	0.0807	0.2460	0.9284	0.9284	0.9284	0.9284
11.0	0.0720	0.2548	0.9274	0.9273	0.9274	0.9274
12.0	0.0692	0.2570	0.9276	0.9275	0.9276	0.9276
13.0	0.0608	0.2612	0.9274	0.9273	0.9274	0.9274
13.0	0.0608	0.2460	0.9284	0.9284	0.9284	0.9284

Evaluation Results

Metric	Score
Accuracy	0.9284
F1 (weighted)	0.9284
Precision (weighted)	0.9284
Recall (weighted)	0.9284

Per-Class Performance

Class	Precision	Recall	F1-Score	Support
Positive	0.9405	0.9458	0.9432	2105
Neutral	0.9110	0.9090	0.9100	1407
Negative	0.9278	0.9227	0.9252	1643

Confusion Matrix

Usage

Pipeline API

from transformers import pipeline

classifier = pipeline(
    "sentiment-analysis",
    model="will702/indo-roBERTa-financial-sentiment-v2"
)

result = classifier("IHSG ditutup menguat 1.5% didorong aksi beli investor asing")
# {'label': 'LABEL_0', 'score': 0.98}  → Positive

Manual Inference

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model = AutoModelForSequenceClassification.from_pretrained(
    "will702/indo-roBERTa-financial-sentiment-v2"
)
tokenizer = AutoTokenizer.from_pretrained(
    "will702/indo-roBERTa-financial-sentiment-v2"
)

text = "Rupiah melemah tajam terhadap dolar AS akibat sentimen global"
inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True)
outputs = model(**inputs)

label_map = {0: "Positive", 1: "Neutral", 2: "Negative"}
predicted = torch.argmax(outputs.logits, dim=1).item()
print(f"Sentiment: {label_map[predicted]}")

Improvements Over Base Model

Aspect	Base (v1)	This Model (v2)
Datasets	1 (intanm only)	3 (intanm + CNBC + SmSA)
Learning Rate	2e-5	1e-5 (preserves prior knowledge)
Scheduler	Linear	Cosine with warmup
Primary Metric	Accuracy	F1 (weighted)
Early Stopping	patience=2	patience=3

Limitations

• Primarily trained on formal financial news — may underperform on very informal social media text or slang
• Label mapping is non-standard (0=Positive) — ensure downstream systems account for this
• Augmented data includes synthetic samples which may not perfectly reflect real-world distribution

Citation

@misc{indo_roberta_fin_v2_2026,
  title        = {IndoRoBERTa Financial Sentiment v2},
  author       = {Gregorius Willson},
  howpublished = {\url{https://huggingface.co/will702/indo-roBERTa-financial-sentiment-v2}},
  year         = {2026},
  note         = {Fine-tuned from ihsan31415/indo-roBERTa-financial-sentiment with multi-source data and augmentation},
}

Acknowledgments

• Original model by Khoirul Ihsan
• Base architecture: w11wo/indonesian-roberta-base-sentiment-classifier
• CNBC Indonesia dataset by triagungj
• IndoNLU benchmark by indonlp