features/text_classifier/inferencer.py

import torch
from .model_loader import get_model_tokenizer

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

def perplexity_to_ai_likelihood(ppl: float) -> float:
    # You can tune these parameters
    min_ppl = 10     # very confident it's AI
    max_ppl = 100    # very confident it's human

    # Clamp to bounds
    ppl = max(min_ppl, min(ppl, max_ppl))

    # Invert and scale: lower perplexity -> higher AI-likelihood
    likelihood = 1 - ((ppl - min_ppl) / (max_ppl - min_ppl))

    return round(likelihood * 100, 2)


def classify_text(text: str):
    model, tokenizer = get_model_tokenizer()
    inputs = tokenizer(text, return_tensors="pt",
                       truncation=True, padding=True)
    input_ids = inputs["input_ids"].to(device)
    attention_mask = inputs["attention_mask"].to(device)

    with torch.no_grad():
        outputs = model(
            input_ids, attention_mask=attention_mask, labels=input_ids)
        loss = outputs.loss
        perplexity = torch.exp(loss).item()

    if perplexity < 55:
        result = "AI-generated"
    elif perplexity < 80:
        result = "Probably AI-generated"
    else:
        result = "Human-written"
    likelihood_result=perplexity_to_ai_likelihood(perplexity)
    return result, perplexity,likelihood_result