README.md

---
language: vi
license: mit
tags:
  - text-classification
  - sentiment-analysis
  - vietnamese
  - finance
  - phobert
base_model: vinai/phobert-base-v2
---

# FiNTA Sentiment Model

**Task:** Sentiment Analysis — Vietnamese Financial News
**Base model:** `vinai/phobert-base-v2`
**Accuracy:** N/A | **Macro F1:** N/A

## Labels (sentiment chính 3-class)

| ID | Label | Mô tả | Score range |
|----|-------|-------|-------------|
| 0 | Bullish | Lạc quan, tích cực | +0.5 → +1.0 |
| 1 | Bearish | Bi quan, tiêu cực | -1.0 → -0.5 |
| 2 | Uncertainty | Không rõ xu hướng | -0.2 → +0.2 |

Raw LLM labels are merged for the primary model:

- Hype/FOMO -> Bullish
- Panic -> Bearish
- Contrarian -> Uncertainty

## Usage

```python
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

tokenizer = AutoTokenizer.from_pretrained("Marky12345/BigData_2026")
model = AutoModelForSequenceClassification.from_pretrained("Marky12345/BigData_2026")

text = "Cổ phiếu VNM tăng mạnh, nhà đầu tư lạc quan về triển vọng quý 4"
inputs = tokenizer(text, return_tensors="pt", max_length=256,
                   truncation=True, padding="max_length")

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

labels = ["Bullish", "Bearish", "Uncertainty"]
print(labels[pred_id])
```

## Training

Fine-tuned on Vietnamese financial news articles labeled by LLM silver labels.
The model predicts the primary 3-class sentiment. Rare market-tone labels are
kept in the dataset for audit and should be handled as separate tags when enough
examples are available.