phobert-sentiment-vn-finance

---

license: apache-2.0 language:

• vi tags:
• text-classification
• sentiment-analysis
• vietnamese
• finance
• stock-market
• phobert
• lora
• peft base_model: vinai/phobert-base-v2 pipeline_tag: text-classification library_name: peft metrics:
• f1
• accuracy
• precision
• recall

---

PhoBERT Vietnamese Financial Sentiment (LoRA, v0)

A LoRA-adapted PhoBERT-base-v2 fine-tuned for 3-class sentiment classification on Vietnamese financial news (stock market, listed companies, macro news).

This is a proof-of-concept (v0) trained on a tiny hand-labeled dataset of 60 sentences as part of a 14-day learning sprint. A v1 with a 500-sentence augmented dataset is planned.

🇻🇳 Đây là phiên bản v0 (warm-up), fine-tune PhoBERT-base-v2 bằng LoRA trên 60 câu tin tài chính tiếng Việt tự gán nhãn để học flow PEFT. Phiên bản v1 với 500 câu augment qua LLM sẽ được release sau.

Labels

How to Use

from peft import PeftModel
from transformers import AutoModelForSequenceClassification, AutoTokenizer, pipeline

REPO_ID = "kienthinh/phobert-sentiment-vn-finance-v0"

base = AutoModelForSequenceClassification.from_pretrained(
    "vinai/phobert-base-v2", num_labels=3
)
model = PeftModel.from_pretrained(base, REPO_ID)
tokenizer = AutoTokenizer.from_pretrained(REPO_ID)

clf = pipeline("text-classification", model=model, tokenizer=tokenizer)

samples = [
    "ACB công bố lợi nhuận quý 3 tăng mạnh, vượt kỳ vọng thị trường.",
    "Cổ phiếu HPG bị bán tháo do tin tiêu cực ngành thép.",
    "VN-Index đóng cửa ở mức 1.250 điểm.",
]
for r, t in zip(clf(samples), samples):
    print(f"{r['label']:10s} ({r['score']:.2f})  ← {t}")

Output (mapping: LABEL_0 = Negative, LABEL_1 = Neutral, LABEL_2 = Positive):
LABEL_2 (0.53)  ← ACB công bố lợi nhuận quý 3 tăng mạnh, vượt kỳ vọng thị trường.
LABEL_0 (0.79)  ← Cổ phiếu HPG bị bán tháo do tin tiêu cực ngành thép.
LABEL_0 (0.40)  ← VN-Index đóng cửa ở mức 1.250 điểm.

Training Data

- Source: 60 sentences hand-labeled by the author covering Vietnamese stock market topics: listed companies (ACB, VCB, FPT, HPG, VNM, MBB, MSN, NVL, MWG, GAS,
 PNJ, ...), macro events, and market index movements.
- Class balance: 20 Positive / 20 Negative / 20 Neutral (perfectly balanced before stratified split).
- Split: 70% train / 15% validation / 15% test, stratified.
- Domain: Vietnamese financial news headlines (informal, short-form).

⚠️ Limitation: 60 sentences is far below production scale. Expect overfitting on phrases that appear in training and weak generalization to unseen patterns.

Training Procedure

Base Model

- vinai/phobert-base-v2 (135M parameters)

LoRA Configuration

┌──────────────────┬──────────────────────────┐
│      Param       │          Value           │
├──────────────────┼──────────────────────────┤
│ r (rank)         │ 8                        │
├──────────────────┼──────────────────────────┤
│ lora_alpha       │ 16                       │
├──────────────────┼──────────────────────────┤
│ lora_dropout     │ 0.1                      │
├──────────────────┼──────────────────────────┤
│ target_modules   │ ["query", "value"]       │
├──────────────────┼──────────────────────────┤
│ Trainable params │ 887,811 (0.65% of total) │
└──────────────────┴──────────────────────────┘

Hyperparameters

┌─────────────────────────┬────────────────────┐
│          Param          │       Value        │
├─────────────────────────┼────────────────────┤
│ Epochs                  │ 20                 │
├─────────────────────────┼────────────────────┤
│ Batch size (train/eval) │ 8 / 8              │
├─────────────────────────┼────────────────────┤
│ Learning rate           │ 5e-4               │
├─────────────────────────┼────────────────────┤
│ Warmup ratio            │ 0.1                │
├─────────────────────────┼────────────────────┤
│ Weight decay            │ 0.01               │
├─────────────────────────┼────────────────────┤
│ Precision               │ bf16               │
├─────────────────────────┼────────────────────┤
│ Optimizer               │ AdamW (HF default) │
├─────────────────────────┼────────────────────┤
│ Best model selector     │ F1 (weighted)      │
└─────────────────────────┴────────────────────┘

Compute

- Hardware: NVIDIA A100 (Google Colab Pro)
- Training time: < 1 minute (tiny dataset)
- Carbon footprint: negligible

Evaluation

Evaluated on the held-out validation split (9 sentences, stratified across 3 classes).

┌──────────────────────┬────────┐
│        Metric        │ Score  │
├──────────────────────┼────────┤
│ Accuracy             │ 0.8889 │
├──────────────────────┼────────┤
│ Precision (weighted) │ 0.9167 │
├──────────────────────┼────────┤
│ Recall (weighted)    │ 0.8889 │
├──────────────────────┼────────┤
│ F1 (weighted)        │ 0.8857 │
├──────────────────────┼────────┤
│ Eval loss            │ 0.6396 │
└──────────────────────┴────────┘

Sample Inference (5 unseen sentences)

┌─────────────────────────────────────────────────────────────────────┬───────────┬────────────┬────────────────────────┐
│                              Sentence                               │ Predicted │ Confidence │        Correct?        │
├─────────────────────────────────────────────────────────────────────┼───────────┼────────────┼────────────────────────┤
│ ACB công bố lợi nhuận quý 3 tăng mạnh, vượt kỳ vọng thị trường.     │ Positive  │ 0.53       │ ✅                     │
├─────────────────────────────────────────────────────────────────────┼───────────┼────────────┼────────────────────────┤
│ Thị trường chứng khoán giảm điểm phiên thứ ba liên tiếp.            │ Negative  │ 0.79       │ ✅                     │
├─────────────────────────────────────────────────────────────────────┼───────────┼────────────┼────────────────────────┤
│ VN-Index đóng cửa ở mức 1.250 điểm.                                 │ Negative  │ 0.40       │ ❌ (should be Neutral) │
├─────────────────────────────────────────────────────────────────────┼───────────┼────────────┼────────────────────────┤
│ Cổ phiếu HPG bị bán tháo do tin tiêu cực ngành thép.                │ Negative  │ 0.79       │ ✅                     │
├─────────────────────────────────────────────────────────────────────┼───────────┼────────────┼────────────────────────┤
│ FPT ký hợp đồng lớn với khách hàng Mỹ, triển vọng dài hạn tích cực. │ Positive  │ 0.54       │ ✅                     │
└─────────────────────────────────────────────────────────────────────┴───────────┴────────────┴────────────────────────┘

4 / 5 correct. The misclassified neutral example shows confidence 0.40 — close to the random uniform baseline (0.33), indicating model uncertainty. In
production, a confidence threshold (e.g. 0.5) is recommended to flag low-confidence predictions for human review.

Bias, Risks & Limitations

1. Tiny training set (60 examples) — generalization beyond seen lexical patterns is limited. The model may bias toward Negative when encountering generic
"VN-Index ..." phrasing because half of the negative training examples mention indices.
2. No word segmentation — PhoBERT is normally fed VnCoreNLP-segmented text. This v0 uses raw whitespace tokenization, which is suboptimal.
3. Domain narrow — trained only on stock market / listed companies. May fail on other financial sub-domains (banking ops, crypto, real estate financing).
4. Hand-labeled by single annotator — labels reflect one person's judgment; no inter-annotator agreement check.
5. Not investment advice — outputs must NOT be used as a basis for trading decisions.

Recommended Use

- ✅ Educational / research demos of LoRA + PhoBERT pipeline
- ✅ Starting point for further fine-tuning with larger datasets
- ❌ Production financial sentiment scoring
- ❌ Automated trading signals

Roadmap

- v1 (planned): Train on 500 LLM-augmented Vietnamese financial sentences with VnCoreNLP word segmentation. Target F1 ≥ 0.85 on a separate held-out test set.
- v2 (planned): Multi-aspect sentiment (per-ticker mention) + integration into a real-time news pipeline.

Citation

If you use this model, please cite the base PhoBERT paper:

@inproceedings{phobert,
    title     = {{PhoBERT: Pre-trained language models for Vietnamese}},
    author    = {Dat Quoc Nguyen and Anh Tuan Nguyen},
    booktitle = {Findings of EMNLP},
    year      = {2020}
}

Author

- Trained by: Kien  — Applied Mathematics @ HCMUS
- Context: Day 2 of a 14-day sprint preparing for an AI Engineer application at ACBS (Asia Commercial Bank Securities).

License

Apache 2.0 (inherits from PhoBERT-base-v2).

---