This model is a fine-tuned version of the xlm-roberta-base model, specifically adapted for Natural Language Inference (NLI) tasks in Vietnamese. It classifies the relationship between a given premise and hypothesis into one of three categories: entailment, neutral, or contradiction.

Model and Training

The core of this model is XLM-RoBERTa, a multilingual masked language model pre-trained on text from 100 different languages. This powerful base is then fine-tuned on a Vietnamese NLI dataset. The fine-tuning process adapts the model to the specific nuances of Vietnamese language and the logical reasoning required for NLI.

The training was conducted for 4 epochs with a batch size of 16 for training and 32 for evaluation. It utilized the AdamW optimizer with a learning rate of 2e-5 and a linear scheduler with a 6% warmup. To accelerate the training, FP16 precision was used on a GPU P100.

Dataset

The model was trained and evaluated on a custom Vietnamese NLI dataset with the following splits:

Train: /kaggle/input/train-nli/train.json
Validation (dev): /kaggle/input/train-nli/vinli_dev.json
Test: /kaggle/input/train-nli/vinli_test.json

Each entry in the dataset consists of a premise, a hypothesis, and a corresponding label indicating their relationship.

Performance

The model demonstrates solid performance on the Vietnamese NLI task, achieving the following results:

Validation:

Accuracy: 72.55%
F1-macro: 72.54%

Test:

Accuracy: 73.10%
F1-macro: 73.00%

Classification Report (Test)

Label	Precision	Recall	F1-score	Support
contradiction	0.7061	0.6716	0.6885	737
neutral	0.7240	0.7799	0.7509	777
entailment	0.7631	0.7387	0.7507	750
Accuracy			0.7310	2264
Macro avg	0.7311	0.7301	0.7300	2264

Confusion Matrix (Test)

	Predicted C	Predicted N	Predicted E
Gold C	495	132	110
Gold N	109	606	62
Gold E	97	99	554

Usage

To use this model for your own Vietnamese NLI tasks, you can load it directly from the Hugging Face Hub using the transformers library.

python

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model_name = "lyle49/xlmr-vinli-finetune"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

premise = "Một sào lúa non nuôi con nửa ngày"
hypothesis = "Lúa non chỉ nuôi được nửa ngày"
inputs = tokenizer(premise, hypothesis, return_tensors="pt", truncation=True, max_length=256)

with torch.no_grad():
    logits = model(**inputs).logits
    pred_id = logits.argmax(dim=-1).item()

id2label = model.config.id2label
print("Prediction:", id2label[pred_id])

Notes

This model shows good performance on Vietnamese NLI tasks, including those that involve figurative or idiomatic reasoning.
It can serve as a strong starting point for developing multilingual entailment models. The base checkpoint can be switched to xlm-roberta-large for potentially better performance, though this may require a smaller batch size to accommodate the larger model.

Citation

If you use this model in your research, please use the following citation:

bibtex

@misc{xlmr_vinli_finetune,
  author = {Lê Lý},
  title = {XLM-RoBERTa fine-tuned on Vietnamese NLI},
  year = {2025},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/lyle49/xlmr-vinli-finetune}}
}