app.py

import torch
import gradio as gr
import numpy as np
from transformers import AutoTokenizer, AutoModelForSequenceClassification
from peft import PeftModel

base = "dleemiller/ModernCE-large-nli"
adapter = "modernce-adapter"

tokenizer = AutoTokenizer.from_pretrained(base)

base_model = AutoModelForSequenceClassification.from_pretrained(
    base,
    num_labels=2,
    problem_type="multi_label_classification",
    ignore_mismatched_sizes=True
)

model = PeftModel.from_pretrained(base_model, adapter)
model.eval()

sigmoid = lambda x: 1 / (1 + np.exp(-x))

def predict(query, context, output):
    sep = tokenizer.sep_token
    inp = sep.join([query, context, output])

    inputs = tokenizer(
        inp,
        return_tensors="pt",
        truncation=True
    )

    with torch.no_grad():
        outputs = model(**inputs)
        logits = outputs.logits.cpu().numpy()[0]

    probs = sigmoid(logits)

    labels = {
        "evident_conflict": float(probs[0]),
        "baseless_info": float(probs[1])
    }

    hallucinated = any(p > 0.5 for p in probs)

    return {
        "hallucinated": hallucinated,
        "probabilities": labels
    }

gr.Interface(
    fn=predict,
    inputs=[
        gr.Textbox(label="Query"),
        gr.Textbox(label="Context"),
        gr.Textbox(label="Model Output")
    ],
    outputs="json",
    title="ModernCE Hallucination Detector",
    description="Detects hallucinations using ModernCE + LoRA adapter"
).launch()