Models & Checkpoints¶
VietQuill provides officially trained checkpoints published on the Hugging Face Hub under the ngwgsang organization.
Official Model Hub Collection¶
VietQuill models are divided into two series based on training data characteristics:
1. Tsubaki Series¶
The Tsubaki series models are trained on public research datasets (ViSP for sentences and ViQP for questions). Ideal for standard sentence rewriting, research benchmarks, and question variations:
* The ensemble Hub repository bundles both sentence and question subfolders.
2. Ume Series¶
The Ume series models are trained on 100K synthesized data comprising longer, more structurally complex, and diverse sentence pairs (sentences: ngwgsang/vietquill-qcpg-100k-synthesis-sentence, questions: ngwgsang/vietquill-qcpg-100k-synthesis-question):
* The ensemble Hub repository bundles both sentence and question subfolders.
Model Architecture Details¶
1. Paraphrase Generator (ViT5)¶
Both Tsubaki (vietquill-vit5-base-tsubaki) and Ume (vietquill-vit5-base-ume) share the exact same ViT5 architecture, input representation, and subfolder layout, differing only in the training data:
- Base Architecture: ViT5 (Vietnamese T5 pre-trained on large-scale Vietnamese corpus).
- Subfolder Structure:
sentence/: Fine-tuned for sentence rewriting (Tsubaki: ViSP dataset; Ume:vietquill-qcpg-100k-synthesis-sentencewith longer, more complex sentences).question/: Fine-tuned for question rewriting (Tsubaki: ViQP dataset; Ume:vietquill-qcpg-100k-synthesis-question).
- Input Representation:
2. Paraphrase Quality Estimator (vELECTRA)¶
- Model:
vietquill-velectra-estimator-tsubaki - Base Architecture: vELECTRA-base.
- Output Heads: Regression heads predicting:
lexical_score: Measures token substitution / diversity.syntactic_score: Measures syntactic transformation divergence.semantic_score: Measures meaning preservation fidelity.