my_indo_qa_model

This model is a fine-tuned version of google-bert/bert-base-multilingual-uncased on an Indonesian Question Answering dataset.

It achieves the following results on the evaluation set:

• Validation Loss: 1.5991
• F1 Score: 0.0873
• Precision: 0.0920
• Recall: 0.0831

---

Model description

This is a fine-tuned BERT multilingual uncased model for the task of extractive question answering in the Indonesian language. It takes a question and a context as input, and predicts the start and end position of the answer span within the context.

---

Intended uses & limitations

Intended Uses:

• Answering factual questions from Indonesian texts
• Educational and research purposes in NLP or linguistics

Limitations:

• Low F1 score (0.0873) indicates the model may struggle with generalization
• Performance may degrade significantly with out-of-domain or noisy texts
• Not recommended for production use without further fine-tuning

---

Training and evaluation data

The model was fine-tuned using a custom Indonesian QA dataset consisting of:

• 2,551 training samples
• 638 test samples

The dataset is structured in SQuAD format with context, question, and answer fields.

---

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• Learning rate: 2e-05
• Train batch size: 16
• Eval batch size: 16
• Epochs: 3
• Optimizer: AdamW (betas=(0.9, 0.999), epsilon=1e-08)
• Scheduler: Linear
• Seed: 42

---

Training results

Training Loss	Epoch	Step	Validation Loss
No log	1.0	160	1.8313
No log	2.0	320	1.5645
No log	3.0	480	1.5991

---

Evaluation metrics

Metric	Value
F1 Score	0.0873
Precision	0.0920
Recall	0.0831

---

Framework versions

• Transformers: 4.54.0
• PyTorch: 2.6.0+cu124
• Datasets: 4.0.0
• Tokenizers: 0.21.2

---