Dataset size: 16,175 examples.
Shared fixed splits for S and T: 12,940 train / 1,617 development / 1,618 test.
Split seed: 42; jointly stratified by sentiment–topic labels with exact duplicate groups kept in one split.
Fine-tuning seeds: [42, 22, 202].
Training: 3 epoch(s), AdamW, learning rate 2e-05, weight decay 0.01, warmup ratio 0.1.
Effective train batch size: 8.
Maximum sequence length: 256.
Input mode: raw normalized Vietnamese text.
No class weighting, resampling, external metadata, images, engagement features, or test-time model selection.
Checkpoints are selected by development Macro-F1. The representative published checkpoint is seed 42, selected only by development Macro-F1.

Results

Test metrics are reported as mean ± sample standard deviation over seeds [42, 22, 202].

Metric	Mean ± std
Test Macro-F1	0.8144 ± 0.0215
Test accuracy	0.9483 ± 0.0029
Test macro precision	0.8832 ± 0.0033
Test macro recall	0.7821 ± 0.0245
Development Macro-F1	0.8346 ± 0.0147

Per-seed results

seed	dev_macro_f1	test_macro_f1	test_accuracy	micro_batch_size	gradient_accumulation_steps
22.000000	0.839254	0.822235	0.950556	8.000000	1.000000
42.000000	0.846464	0.830901	0.949320	8.000000	1.000000
202.000000	0.818184	0.790007	0.944994	8.000000	1.000000

Label mapping

json

{
  "0": "negative",
  "1": "neutral",
  "2": "positive"
}

Usage

python

import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer

model_id = "BaoNhan/xlm-roberta-base-UITVSFC-S"
tokenizer = AutoTokenizer.from_pretrained(model_id, use_fast=False)
model = AutoModelForSequenceClassification.from_pretrained(model_id)

text = "Đây là nội dung tiếng Việt cần phân loại."
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=256)
with torch.no_grad():
    probabilities = model(**inputs).logits.softmax(dim=-1)[0]
predicted_id = int(probabilities.argmax())
print(model.config.id2label[predicted_id], probabilities.tolist())

Files

aggregate_metrics.json: aggregate metrics and training manifest.
artifacts/per_seed_results.csv: one row per fine-tuning seed.
artifacts/seed_*_confusion_matrix.csv: confusion matrix for each seed.
artifacts/seed_*_classification_report.json: per-class metrics.
artifacts/seed_*_test_predictions.csv: IDs, gold/predicted labels and probabilities; raw text is excluded.

Limitations

UIT-VSFC contains student feedback from a specific educational context. Performance may not transfer to other institutions, domains, informal writing styles, or newly emerging vocabulary. Predictions should not be treated as explanations of student intent.

Dataset citation

bibtex

@InProceedings{8573337,
  author={Nguyen, Kiet Van and Nguyen, Vu Duc and Nguyen, Phu X. V. and Truong, Tham T. H. and Nguyen, Ngan Luu-Thuy},
  booktitle={2018 10th International Conference on Knowledge and Systems Engineering (KSE)},
  title={UIT-VSFC: Vietnamese Students' Feedback Corpus for Sentiment Analysis},
  year={2018},
  pages={19--24},
  doi={10.1109/KSE.2018.8573337}
}