Vietnamese extractive sentence selectors advanced ML v2
Logistic Regression and MLP with StandardScaler, fitted on oracle labels from 3,000 train articles. Hyperparameters selected using 500 validation articles. All three replication seeds (42, 123, 2026) are provided; default demonstration seed 42 was fixed before test, not selected using test scores.
Use bkai-foundation-models/vietnamese-bi-encoder, revision 84f9d9ada0d1a3c37557398b9ae9fcedcdf40be0, frozen and L2 normalized. Do not substitute the fine-tuned encoder at kttt294/vietnamese-sbert-advanced-ml-v2: these heads were trained on pretrained features. Vietnamese word segmentation uses underthesea.
Each joblib artifact contains model, scaler, params, seed, training trace and timing. Feature order is defined in src/advanced_ml.py::features; prediction uses predict_head. Select exactly 3 highest-scoring sentences and restore source order. Use the versions in environment.json for reproducibility. Only deserialize these artifacts if you trust their publisher (joblib is pickle based).
Validation-selected settings: Logistic C=0.01, L2; MLP 128 ReLU units, Adam, L2 alpha=0.1, lr=0.001, early stopping by validation log-loss. On 1,000 held-out articles, mean ROUGE-L: Logistic 0.191300, MLP 0.186970, pretrained K-Means 0.176644, Lead-3 0.188761. Lead-3 leads ROUGE-1/2 and BERTScore. These are method comparisons: classifiers additionally use position, length and centroid features; no embedding-only ablation was performed. Do not attribute all gains to architecture. All ROUGE Vietnamese scores use Unicode syllable tokens, not the old ASCII tokenizer.
Companion fine-tuned encoder: https://huggingface.co/kttt294/vietnamese-sbert-advanced-ml-v2 Source: https://github.com/kttt294/Extractive-Summarizer