# Summary Evaluator (ViDeBERTa base + 2-layer MLP)

This repo contains a Vietnamese summary evaluation model:
- **Backbone**: `Fsoft-AIC/videberta-base`
- **Head**: 2-layer MLP for 3 regression scores:
  - faithfulness, coherence, relevance (range depends on training data, commonly 1–5)

## Files
- `config.json`, `model.safetensors`, tokenizer files: backbone encoder
- `regressor.pt`: regression head weights
- `head_config.json`: head meta
- `modeling_summary_evaluator.py`: convenience loader (`from_pretrained_custom`)
- `training_args.json`: training-time hyperparameters

## Usage (inference)
```