AI legal assistant for disabilities

app.py

@@ -0,0 +1,120 @@
+import gradio as gr
+from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, pipeline
+import datetime
+import torch
+from trl import PPOTrainer, PPOConfig, AutoModelForSeq2SeqLMWithValueHead, create_reference_model, set_seed
+from transformers import pipeline
+
+# Load FLAN-T5 for Legal Q&A
+model_name = "google/flan-t5-small"
+tokenizer = AutoTokenizer.from_pretrained(model_name)
+model = AutoModelForSeq2SeqLMWithValueHead.from_pretrained(model_name)
+
+# Create a reference model for PPO
+ref_model = create_reference_model(model)
+
+# PPO Configuration
+config = PPOConfig(
+    batch_size=1,
+    learning_rate=1e-5,
+    mini_batch_size=1,
+    ppo_epochs=1  # Minimal epochs
+)
+
+# Create PPO Trainer
+ppo_trainer = PPOTrainer(
+    config=config,
+    model=model,
+    ref_model=ref_model,
+    tokenizer=tokenizer
+)
+
+# Translation Models (English ↔ Hindi)
+translator_en_hi = AutoModelForSeq2SeqLM.from_pretrained("Helsinki-NLP/opus-mt-en-hi")
+tokenizer_en_hi = AutoTokenizer.from_pretrained("Helsinki-NLP/opus-mt-en-hi")
+
+translator_hi_en = AutoModelForSeq2SeqLM.from_pretrained("Helsinki-NLP/opus-mt-hi-en")
+tokenizer_hi_en = AutoTokenizer.from_pretrained("Helsinki-NLP/opus-mt-hi-en")
+
+# Translation Function
+def translate(text, src_lang, tgt_lang):
+    if src_lang == "English" and tgt_lang == "Hindi":
+        inputs = tokenizer_en_hi(text, return_tensors="pt", padding=True, truncation=True)
+        outputs = translator_en_hi.generate(**inputs)
+        return tokenizer_en_hi.decode(outputs[0], skip_special_tokens=True)
+    elif src_lang == "Hindi" and tgt_lang == "English":
+        inputs = tokenizer_hi_en(text, return_tensors="pt", padding=True, truncation=True)
+        outputs = translator_hi_en.generate(**inputs)
+        return tokenizer_hi_en.decode(outputs[0], skip_special_tokens=True)
+    else:
+        return "Translation for this pair not supported yet!"
+
+# Generate Complaint Template
+def generate_complaint(issue):
+    date = datetime.datetime.now().strftime("%d-%m-%Y")
+    template = f"""
+[Your Name]  
+[Your Address]  
+{date}  
+
+To Whom It May Concern,  
+
+**Subject: Complaint Regarding {issue}**  
+
+I am writing to formally lodge a complaint regarding {issue}. The incident occurred on [Date/Location]. The specific details are as follows:  
+
+- Issue: {issue}  
+- Evidence: [Provide Evidence]  
+
+I kindly request you to take appropriate action as per the legal guidelines.  
+
+Yours sincerely,  
+[Your Name]  
+"""
+    return template.strip()
+
+# Handle Legal Q&A with PPO
+def handle_legal_query(query, language):
+    if language != "English":
+        query = translate(query, language, "English")
+
+    # Tokenize input
+    inputs = tokenizer(query, return_tensors="pt", padding=True, truncation=True)
+
+    # Generate Response
+    outputs = model.generate(**inputs, max_length=150)
+    response = tokenizer.decode(outputs[0], skip_special_tokens=True)
+
+    # Reward Signal for PPO (basic reward)
+    reward = torch.tensor([1.0]) if "legal" in response.lower() else torch.tensor([-1.0])
+
+    # PPO Step (Reinforcement Learning)
+    ppo_trainer.step([query], [outputs], [reward])
+
+    if language != "English":
+        response = translate(response, "English", language)
+
+    return response
+
+# Define Gradio Interface
+with gr.Blocks(css=".container {width: 100%; max-width: 600px;}") as app:
+    gr.Markdown("# AI Legal Assistant\n### Ask legal questions and generate complaints")
+
+    with gr.Row():
+        query = gr.Textbox(label="Ask your legal question", placeholder="What are my rights as a disabled person?")
+        lang = gr.Dropdown(["English", "Hindi"], label="Language", value="English")
+
+    with gr.Row():
+        submit_btn = gr.Button("Get Legal Advice")
+        output = gr.Textbox(label="Legal Advice", placeholder="Legal advice will appear here")
+
+    with gr.Row():
+        issue = gr.Textbox(label="Describe your issue", placeholder="Facing discrimination at work...")
+        generate_btn = gr.Button("Generate Complaint")
+        complaint_output = gr.Textbox(label="Generated Complaint", placeholder="Complaint template will appear here")
+
+    submit_btn.click(handle_legal_query, inputs=[query, lang], outputs=output)
+    generate_btn.click(generate_complaint, inputs=issue, outputs=complaint_output)
+
+# Launch the app on Hugging Face free tier
+app.launch()