Model Card: finbeto-lora

Purpose

finbeto-lora analyzes sentiment in Spanish financial news headlines. It is designed for financial text classification (positive, negative, neutral) in Spanish.

Training Details

• Base model: dccuchile/bert-base-spanish-wwm-cased
• Adapter: LoRA (PEFT)
• Datasets:
  • data/raw/financial_news.csv (Spanish headlines)
  • data/processed/financial_phrasebank_google_translate_es.csv (PhraseBank, translated)
• Key hyperparameters:
  • learning_rate: ~2.8e-5
  • weight_decay: 0.1
  • num_train_epochs: 3
  • batch_size: 16 (train), 32 (eval)
  • LoRA rank: 4, alpha: 32, dropout: 0.1
• Precision: fp16

Metrics

• Classification Report:

---

	precision	recall	f1-score	support
Positive	0.78	0.69	0.73	1095
Negative	0.73	0.82	0.77	898
Neutral	0.78	0.81	0.80	750
accuracy			0.76	2743
macro avg	0.77	0.77	0.77	2743
weighted avg	0.77	0.76	0.76	2743

---

Usage Example

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

tokenizer = AutoTokenizer.from_pretrained("EuclidesHernandez/finbeto")
model = AutoModelForSequenceClassification.from_pretrained("EuclidesHernandez/finbeto")

text = "La empresa reportó un crecimiento significativo en el último trimestre."
inputs = tokenizer(text, return_tensors="pt")
with torch.no_grad():
        logits = model(**inputs).logits
        pred = torch.argmax(logits, dim=1).item()
        print(["negative", "neutral", "positive"][pred])

Contact:

For more information or to stay in touch, please visit: https://github.com/euclideshh/FinancialNewsSentimentAnalysis

License

MIT License