hadimaster65555/master-indonesian-decision-e3

Master - Experimental Decision Model

An experimental question-conditioned decision-scoring model trained using mmBERT-base.

Training datasets

• IndoNLU SmSA
• IndoNLU EmoT
• IndoNLU CASA
• IndoNLU HoASA
• IndoNLU WReTE

Optional additional dataset:

• IndoNLI, when enabled in the experiment

Training configuration

Experiment: E3

Architecture:

• Pretrained mmBERT-base encoder
• Shared candidate-scoring head
• Partial fine-tuning, full fine-tuning, or LoRA, depending on experiment settings

Evaluation

See test_metrics.csv and aspect_metrics.csv for actual results from this experiment.

Loading instructions

This repository contains a custom PyTorch decision scorer.

It is NOT directly loadable with the default Hugging Face text-classification pipeline.

To restore:

1. Load jhu-clsp/mmBERT-base.
2. Recreate DecisionScorer from the training notebook.
3. Load best_trainable_weights.pt.
4. Restore the tokenizer and calibration settings.

Limitations

This is a research prototype.

Performance on unseen decision domains, new candidate descriptions, and high-stakes applications is not established.

Dataset licenses and upstream terms must be reviewed before commercial redistribution.