app.py

import gradio as gr
from transformers import AutoTokenizer, AutoModelForSequenceClassification
from sentence_transformers import SentenceTransformer, util
import torch
import torch.nn.functional as F

# Load SBERT model
sbert_model = SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2")

# Load NLI model
nli_model_name = "tasksource/ModernBERT-base-nli"
nli_tokenizer = AutoTokenizer.from_pretrained(nli_model_name)
nli_model = AutoModelForSequenceClassification.from_pretrained(nli_model_name)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
nli_model.to(device)

# SBERT function
def compute_similarity(text1, text2):
    embeddings = sbert_model.encode([text1, text2], convert_to_tensor=True)
    similarity = float(util.pytorch_cos_sim(embeddings[0], embeddings[1])[0])
    
    interpretation = ""
    if similarity > 0.9:
        interpretation = "🟢 Very High Similarity"
    elif similarity > 0.75:
        interpretation = "🟡 Moderate Similarity"
    elif similarity > 0.5:
        interpretation = "🟠 Low Similarity"
    else:
        interpretation = "🔴 Very Low Similarity"
    
    return round(similarity, 4), interpretation

# NLI function
def check_entail(premise, hypothesis):
    inputs = nli_tokenizer(premise, hypothesis, return_tensors="pt", truncation=True, padding=True, max_length=512).to(device)
    with torch.no_grad():
        logits = nli_model(**inputs).logits
    probs = torch.softmax(logits, dim=-1)[0]
    label = ["entailment", "neutral", "contradiction"][torch.argmax(probs).item()]
    return label, { "entailment": float(probs[0]), "neutral": float(probs[1]), "contradiction": float(probs[2]) }

def run_bi_direction(a, b):
    res1 = check_entail(a, b)
    res2 = check_entail(b, a)
    return res1[0], res1[1], res2[0], res2[1]

# Build the interface
with gr.Blocks() as demo:
    gr.Markdown("# ✍️ Essay Comparison Tool")

    with gr.Tab("Semantic Similarity (SBERT)"):
        a1 = gr.Textbox(label="Essay A", lines=8)
        b1 = gr.Textbox(label="Essay B", lines=8)
        sim_button = gr.Button("Compare Similarity")
        sim_score = gr.Textbox(label="Cosine Similarity (0–1)")
        sim_interpret = gr.Textbox(label="Interpretation")
        sim_button.click(fn=compute_similarity, inputs=[a1, b1], outputs=[sim_score, sim_interpret])

    with gr.Tab("Bidirectional Entailment (NLI)"):
        a2 = gr.Textbox(label="Essay A (Original)", lines=8)
        b2 = gr.Textbox(label="Essay B (Modified)", lines=8)
        nli_button = gr.Button("Run Entailment Check")
        ab_label = gr.Textbox(label="A ➜ B Label")
        ab_scores = gr.JSON(label="A ➜ B Scores")
        ba_label = gr.Textbox(label="B ➜ A Label")
        ba_scores = gr.JSON(label="B ➜ A Scores")
        nli_button.click(fn=run_bi_direction, inputs=[a2, b2], outputs=[ab_label, ab_scores, ba_label, ba_scores])

if __name__ == "__main__":
    demo.launch()