README.md

---
license: apache-2.0
tags:
- evaluation
- summarization
- vietnamese
- reward-model
- full-document
---

# MultiEvalSumVietN (Full-Document Evaluator)

Vietnamese summarization evaluator trained on (doc, summary) pairs.
Outputs 3 scores in **[0, 1]**:
- Faithfulness
- Coherence
- Relevance

## Files
- `config.json` + `model.safetensors`: backbone (Transformers compatible)
- tokenizer files (copied from `Fsoft-AIC/videberta-base` if available)
- `arch_config.json`: full-doc hyperparams
- `trunk.pt`, `head_*.pt`: regression heads
- `seg_attn.pt`: segment attention module (when `agg_type="attn"`)
- `modeling_summary_evaluator.py`: reference loader
- `package_versions.json`: environment snapshot

## Quick load (reference)
```python
from huggingface_hub import snapshot_download
from modeling_summary_evaluator import FullDocEvaluator

base_dir = snapshot_download("phuongntc/Multi_EvalSumVietN_FullDoc")
model = FullDocEvaluator(base_dir).eval()