squadgoals404/XLM-Prohori-v2

🌐 XLM-Prohori-v2: Bangla/English SMS Smishing Classifier

Repository: squadgoals404/XLM-Prohori-v2
Base Model: xlm-roberta-base

---

📌 Overview

XLM-Prohori-v2 is a fine-tuned XLM-RoBERTa-base model for detecting smishing (SMS phishing) in Bangla and English.
It classifies SMS into three categories:

• normal → Casual, harmless, informational texts
• promo → Promotional/advertising messages
• smish → Smishing (phishing via SMS) attempts

---

📊 Dataset

• Total samples (after deduplication): ~4,507
• Languages: Bangla, English, Banglish
• Labels: balanced across normal, promo, smish
• Preprocessing: All URLs normalized to [LINK]; duplicates removed; stratified train/val/test split
• Splits: Train=3064, Val=541, Test=902 (verified zero overlap)

The raw dataset is not publicly released for privacy reasons. Some synthetic smish examples were included to balance classes.

---

📈 Performance

• Validation Accuracy: ~97.60%
• Test Accuracy: ~97.23%

Confusion matrices indicate generally balanced performance, with minor confusion between promo and smish in link-heavy texts.

---

🚀 Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch.nn.functional as F
import torch

model_id = "squadgoals404/XLM-Prohori-v2"
tok = AutoTokenizer.from_pretrained(model_id)
mdl = AutoModelForSequenceClassification.from_pretrained(model_id)

# Force semantic class names in-memory
CLASS_NAMES = ["normal", "promo", "smish"]  # make sure the order matches 0,1,2
mdl.config.id2label = {i: c for i, c in enumerate(CLASS_NAMES)}
mdl.config.label2id = {c: i for i, c in enumerate(CLASS_NAMES)}

text = "Bank Account temporarily locked—identity verify করতে জরুরি কল 017XX-XXXXXX"
inputs = tok(text, return_tensors="pt")
with torch.no_grad():
    probs = F.softmax(mdl(**inputs).logits, dim=-1).squeeze().tolist()

print({CLASS_NAMES[i]: round(p, 4) for i, p in enumerate(probs)})