PhoBERT-VFQT-Cls

Vietnamese financial question-type classifier fine-tuned from vinai/phobert-base-v2.

Model Details

Model Description

• Model type: Encoder-only Transformer (PhoBERT / RoBERTa) for sequence classification (RobertaForSequenceClassification)
• Task: Financial question type classification (4 labels)
• Language: Vietnamese
• Finetuned from: vinai/phobert-base-v2
• Pipeline tag: text-classification
• Developed by: tiam4tt

The training notebook uses AutoTokenizer, AutoModelForSequenceClassification, and a custom WeightedTrainer(Trainer) that applies class-weighted cross entropy.

Labels

The label mapping used during training (saved in config.json) is:

id	label
0	COMPARISON
1	FACTOID
2	SUMMARY
3	VERIFICATION

Model Sources

• Base model: https://huggingface.co/vinai/phobert-base-v2

Uses

Direct Use

• Route Vietnamese financial questions to the right downstream module (factoid retrieval vs. summarization vs. verification vs. comparison).
• Dataset labeling / analysis for question-type distribution.

Out-of-Scope Use

• Non-Vietnamese text or non-question inputs.
• High-stakes domains (medical/legal) without additional validation.

Bias, Risks, and Limitations

• Domain shift: trained on a specific finance/news-style dataset; performance may drop on other writing styles or topics.
• Single-label constraint: real questions can be multi-intent; the model outputs exactly one label.
• Downstream risk: misclassification can route requests into the wrong pipeline.

Usage (Run Example)

Install deps:

pip install -U "transformers" "torch" "sentencepiece"

Run a single prediction:

import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification

MODEL_ID = "tiam4tt/PhoBERT-VFQT-Cls"  # or a local path to this folder

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForSequenceClassification.from_pretrained(MODEL_ID)
model.eval()

text = "Tăng trưởng GDP quý này so với quý trước như thế nào?"

inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=256)
with torch.no_grad():
    logits = model(**inputs).logits

pred_id = int(torch.argmax(logits, dim=-1).item())
id2label = model.config.id2label
pred_label = id2label.get(pred_id, id2label.get(str(pred_id), f"LABEL_{pred_id}"))

print(pred_label)

# Expects: "COMPARISON"

Evaluation Results

• Accuracy: 0.9486
• Macro F1: 0.9508
• Weighted F1: 0.9485

Per-class report (test set)

label	precision	recall	f1
COMPARISON	0.9576	0.9480	0.9528
FACTOID	0.9153	0.9080	0.9116
SUMMARY	0.9506	0.9620	0.9563
VERIFICATION	0.9784	0.9864	0.9824

Technical Specifications

Compute Infrastructure

• GPU: Tesla T4
• CUDA: 12.8
• PyTorch: 2.10.0+cu128
• Transformers: 5.0.0 (recorded in config.json)

Model Card Contact

• tiam4tt