| 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() | |