lyle49/xlmr-vinli-finetune-full-6k

🧠 XLM-RoBERTa for Vietnamese NLI — Full 6K Finetune

Fine-tuned XLM-RoBERTa-base model for Vietnamese Natural Language Inference (NLI) — trained on a manually curated and balanced dataset of ~6k premise–hypothesis pairs. The model predicts one of three labels:

• c → Contradiction
• n → Neutral
• e → Entailment

---

🚀 Overview

Item	Description
Base model	xlm-roberta-base
Framework	PyTorch + 🤗 Transformers (v4.43.3)
Dataset size	14,116 train / 1,765 val / 1,765 test
Sequence length	256
Task	Vietnamese NLI (Entailment / Contradiction / Neutral)
Language	Vietnamese 🇻🇳
License	MIT (default HF sharing)
Author	@lyle49

---

📚 Dataset

The dataset (full_data_true_conflict.json) includes Vietnamese NLI pairs with carefully normalized labels:

Split	Size	Label distribution
Train	14,116	c: 4,705 / n: 4,706 / e: 4,705
Validation	1,765	c: 589 / n: 588 / e: 588
Test	1,765	c: 588 / n: 588 / e: 589

Each record:

{
  "id": "example_001",
  "premise": "Anh ấy đang chơi bóng đá trên sân.",
  "hypothesis": "Anh ấy đang tham gia một môn thể thao.",
  "label": "e"
}

⚙️ Training Configuration

Setting	Value
Epochs	4
Batch size	8 × 2 (gradient accumulation = 2)
Optimizer	AdamW
Learning rate	2e-5
Warmup ratio	0.06
Weight decay	0.01
Precision	fp16 (AMP)
Gradient checkpointing	✅ Enabled
Label smoothing	0.05
Early stopping	patience = 2
Tokenizer max length	256
Save format	.safetensors

Environment:

• CUDA 12.4
• Transformers 4.43.3
• PyTorch 2.3.0+cu124
• Kaggle P100 GPU

🧩 Model Performance

Split	Accuracy	F1 (macro)	Loss
Validation	0.9807	0.9808	0.2278
Test	0.9796	0.9796	0.2354

Confusion Matrix (Test)

	c	n	e
c	571	9	8
n	4	578	6
e	3	6	580

Classification Report (Test)

Label	Precision	Recall	F1
c	0.9879	0.9711	0.9794
n	0.9747	0.9830	0.9788
e	0.9764	0.9847	0.9806
Macro avg	0.9797	0.9796	0.9796

---

🧪 Usage Example

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model_name = "lyle49/xlmr-vinli-finetune-full-6k"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

premise = "Anh ấy đang chơi bóng đá trên sân."
hypothesis = "Anh ấy đang tham gia một môn thể thao."
inputs = tokenizer(premise, hypothesis, return_tensors="pt", truncation=True, max_length=256)

with torch.no_grad():
    outputs = model(**inputs)
    probs = torch.softmax(outputs.logits, dim=-1)
    label = model.config.id2label[probs.argmax().item()]

print(f"Prediction: {label} ({probs.max().item():.3f})")

Output:

Prediction: e (0.980)

🧭 Notes

• The model achieves ~98% accuracy on a balanced Vietnamese NLI test set.
• It is robust for figurative / paraphrased sentence pairs and works well on both formal and informal text.
• For extended use (e.g., paraphrase detection, contradiction filtering, or multi-lingual entailment), consider fine-tuning with domain data.
• Model is saved as .safetensors for safety and compatibility.

🏷️ Citation

If you use this model in research or product, please cite:

@misc{lyle49_xlmr_vinli_2025,
  author = {Lê, Lý},
  title = {XLM-RoBERTa for Vietnamese NLI — Full 6K Finetune},
  year = {2025},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/lyle49/xlmr-vinli-finetune-full-6k}}
}

🧰 Files in Repository

File	Description
config.json	Model configuration
model.safetensors	Finetuned model weights
tokenizer.json, tokenizer_config.json, special_tokens_map.json, sentencepiece.bpe.model	Tokenizer files
README.md	This documentation

---

⭐ Created by @lyle49 — part of the Vietnamese Figurative NLI Project (2025).