deception-detector

DistilBERT — Deception Risk (fine‑tuned & calibrated)

Short description

• DistilBERT fine‑tuned to predict "deception risk"
• Fine tuned on LIAR, dailiDialog, and general deception datasets
• Probabilities are temperature‑calibrated (see temperature.pt) so you can show a continuous risk score.

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch
import numpy as np

model_dir = "your-username/your-hf-repo"  # or local path to folder with files
device = "cuda" if torch.cuda.is_available() else "cpu"

tok = AutoTokenizer.from_pretrained(model_dir)
model = AutoModelForSequenceClassification.from_pretrained(model_dir).to(device).eval()

# load temperature if present
temp = 1.0
try:
    temp = float(torch.load(f"{model_dir}/temperature.pt", map_location="cpu").get("temperature", 1.0))
except Exception:
    pass

def get_deception_risk(texts):
    if isinstance(texts, str):
        texts = [texts]
    enc = tok(texts, truncation=True, padding=True, return_tensors="pt", max_length=256).to(device)
    with torch.no_grad():
        logits = model(**enc).logits
        probs = torch.softmax(logits / temp, dim=-1)[:, 1].cpu().numpy()  # index 1 = risky class
    return probs

# example
print(get_deception_risk(["This review looks fake", "This is a genuine comment"]))