NLP Coercion Detector — SecureWealth Twin (M3)

Fine-tuned google/muril-base-cased for detecting stress and coercion language in bank AI chat messages across English, Hindi (Devanagari), and Punjabi (Gurmukhi).

Part of the WealthVault's AI system — a bank-grade fraud detection and financial intelligence platform.

---

Model Details

---

Labels

---

Usage

Load and run inference

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model_id  = "NanG01/m3-coercion-bert"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model     = AutoModelForSequenceClassification.from_pretrained(model_id)
model.eval()

def predict_coercion(text: str) -> dict:
    enc = tokenizer(text, return_tensors="pt", truncation=True,
                    padding="max_length", max_length=128)
    with torch.no_grad():
        label = model(**enc).logits.argmax(-1).item()
    return {"chat_stress_language": label, "risk_pts": 20 if label == 1 else 0}

Examples

predict_coercion("They said I must transfer 50000 rupees urgent right now")
# → {"chat_stress_language": 1, "risk_pts": 20}

predict_coercion("Please fast, he told me this is the last warning from income tax department")
# → {"chat_stress_language": 1, "risk_pts": 20}

predict_coercion("Jaldi karo, do minute vich transfer karna hai nahi tan khat khatam ho jaavega")
# → {"chat_stress_language": 1, "risk_pts": 20}

predict_coercion("How can I increase my monthly SIP amount?")
# → {"chat_stress_language": 0, "risk_pts": 0}

predict_coercion("Mera portfolio performance dikhao")
# → {"chat_stress_language": 0, "risk_pts": 0}

---

Output

chat_stress_language feeds directly into M4 Coercion Risk Scorer as one of 12 binary signals (+20 risk pts).

---

Trigger Patterns

Common coercion indicators the model detects:

urgent · hurry · they said · he told me · please fast · last warning · account will be blocked · income tax · jaldi karo · abhi transfer karo

---

Training Data

~490 chat messages across 3 languages:

• 90 English coercion/normal messages
• 400 Hindi coercion/normal messages
• Punjabi messages included

Dataset: SecureWealthTwin_DL_Datasets_v2.xlsx (private)

Regularisation applied:

• Frozen BERT encoder layers 0–9 (only top 2 layers + classifier trained)
• Dropout 0.3 on hidden, attention, and classifier layers
• Weight decay 0.01
• Gradient clipping (max norm 1.0)
• Early stopping (patience=2)

---

---

license: apache-2.0