app.py

import os

# Ensure Java is installed in Hugging Face Space
if not os.path.exists("/usr/bin/java"):
    print("🔄 Installing Java manually...")
    os.system("wget https://download.java.net/openjdk/jdk11/ri/openjdk-11+28_linux-x64_bin.tar.gz")
    os.system("tar -xvzf openjdk-11+28_linux-x64_bin.tar.gz")
    os.system("mv jdk-11 java")
    os.environ["JAVA_HOME"] = os.path.abspath("java")
    os.environ["PATH"] += os.pathsep + os.path.join(os.path.abspath("java"), "bin")

print("✅ Java is installed!")

import language_tool_python

# Initialize LanguageTool
tool = language_tool_python.LanguageTool('en-US')


import gradio as gr
import ftfy
import language_tool_python
import re
import torch
from sentence_transformers import SentenceTransformer, util
from transformers import pipeline, AutoModelForCausalLM, AutoTokenizer


device = "cuda" if torch.cuda.is_available() else "cpu"  # ✅ Use GPU if available

# Load fine-tuned GPT-2 model
model_path = "cpv2280/gpt2-tinystories-generator" # Update if needed
model = AutoModelForCausalLM.from_pretrained(model_path)
tokenizer = AutoTokenizer.from_pretrained(model_path)

# Create a text-generation pipeline
story_generator = pipeline("text-generation", model=model, tokenizer=tokenizer)

# Load NLP tools
tool = language_tool_python.LanguageTool('en-UK')
sentence_model = SentenceTransformer('all-MiniLM-L6-v2')

def refine_story(text):
    """Refines the generated story by fixing encoding, grammar, and redundancy."""
    text = ftfy.fix_text(text)  # Fix encoding
    matches = tool.check(text)  # Check grammar
    text = language_tool_python.utils.correct(text, matches)  # Apply fixes

    # Remove redundant words/phrases
    text = re.sub(r'(\b\w+\b) \1', r'\1', text)  # Remove duplicate words
    text = re.sub(r'(\b\w+ and \w+\b)(,? \1)+', r'\1', text)  # Remove phrase repetitions

    return text

def detect_inconsistencies(text):
    """Checks for logical inconsistencies by comparing sentence similarities."""
    sentences = text.split(". ")
    inconsistencies = []

    # Compare each sentence with the next one
    for i in range(len(sentences) - 1):
        similarity_score = util.pytorch_cos_sim(sentence_model.encode(sentences[i]), sentence_model.encode(sentences[i+1]))

        if similarity_score.item() < 0.3:  # If similarity is low, flag as inconsistent
            inconsistencies.append(f"⚠️ **Possible inconsistency detected:**\n➡ {sentences[i]} \n➡ {sentences[i+1]}")

    return "\n\n".join(inconsistencies) if inconsistencies else "✅ No major inconsistencies detected."

def story_pipeline(prompt):
    """Generates a story, refines it, and checks inconsistencies."""
    # Generate the story
    generated = story_generator(prompt, max_length=200, do_sample=True, temperature=1.0, top_p=0.9, top_k=50, truncation=True)
    raw_story = generated[0]['generated_text']

    # Refine the generated story
    refined_story = refine_story(raw_story)

    # Detect logical inconsistencies
    inconsistencies = detect_inconsistencies(refined_story)

    return raw_story, refined_story, inconsistencies

# ✅ Gradio Interface with Proper Logical Inconsistency Detection
interface = gr.Interface(
    fn=story_pipeline,
    inputs=gr.Textbox(label="Enter Story Prompt", placeholder="Once upon a time..."), 
    outputs=[
        gr.Textbox(label="📖 Generated Story", interactive=True),  # Interactive textbox
        gr.Textbox(label="✅ Refined Story", interactive=True),  # Refined output
        gr.Textbox(label="⚠️ Logical Inconsistencies", interactive=False),  # Shows inconsistencies correctly
    ],
    title="📖 FableWeaver AI",
    description="Generates AI-powered TinyStories using GPT-2 fine-tuned on TinyStories. Automatically refines the story and detects logical inconsistencies."
)

# Launch Gradio app
interface.launch(share="True")