Create app.py

app.py

@@ -0,0 +1,202 @@
+import gradio as gr
+import spaces
+from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
+import torch
+from huggingface_hub import InferenceClient
+import os
+
+# Initialize Cerebras client for Llama 4
+cerebras_client = InferenceClient(
+    "meta-llama/Llama-4-Scout-17B-16E-Instruct",
+    provider="cerebras",
+    token=os.getenv("HF_TOKEN"),
+)
+
+# Global variables for models and tokenizers
+en_es_tokenizer = None
+en_es_model = None
+es_en_tokenizer = None
+es_en_model = None
+
+@spaces.GPU(duration=60)
+def translate_en_to_es(text):
+    global en_es_tokenizer, en_es_model
+    
+    # Initialize EN->ES model if needed
+    if en_es_tokenizer is None or en_es_model is None:
+        en_es_tokenizer = AutoTokenizer.from_pretrained("facebook/nllb-200-distilled-600M", src_lang="eng_Latn", tgt_lang="spa_Latn")
+        en_es_model = AutoModelForSeq2SeqLM.from_pretrained(
+            "facebook/nllb-200-distilled-600M",
+            torch_dtype=torch.float16
+        ).cuda()
+    
+    # Translate
+    inputs = en_es_tokenizer(text, return_tensors="pt", max_length=512, truncation=True).to("cuda")
+    with torch.no_grad():
+        outputs = en_es_model.generate(
+            **inputs,
+            forced_bos_token_id=en_es_tokenizer.convert_tokens_to_ids("spa_Latn"),
+            max_length=512,
+            num_beams=5,
+            early_stopping=True
+        )
+    
+    translation = en_es_tokenizer.decode(outputs[0], skip_special_tokens=True)
+    return translation
+
+@spaces.GPU(duration=60)
+def translate_es_to_en(text):
+    global es_en_tokenizer, es_en_model
+    
+    # Initialize ES->EN model if needed
+    if es_en_tokenizer is None or es_en_model is None:
+        es_en_tokenizer = AutoTokenizer.from_pretrained("facebook/nllb-200-distilled-600M", src_lang="spa_Latn", tgt_lang="eng_Latn")
+        es_en_model = AutoModelForSeq2SeqLM.from_pretrained(
+            "facebook/nllb-200-distilled-600M",
+            torch_dtype=torch.float16
+        ).cuda()
+    
+    # Translate
+    inputs = es_en_tokenizer(text, return_tensors="pt", max_length=512, truncation=True).to("cuda")
+    with torch.no_grad():
+        outputs = es_en_model.generate(
+            **inputs,
+            forced_bos_token_id=es_en_tokenizer.convert_tokens_to_ids("eng_Latn"),
+            max_length=512,
+            num_beams=5,
+            early_stopping=True
+        )
+    
+    translation = es_en_tokenizer.decode(outputs[0], skip_special_tokens=True)
+    return translation
+
+def refine_with_llama(original_text, translation, direction, formality="neutral"):
+    if direction == "en_to_es":
+        refine_prompt = f"""You are an expert Mexican Spanish translator. Refine the following translation and explain your changes:
+
+Original English: {original_text}
+Initial Spanish translation: {translation}
+Formality level: {formality}
+
+Requirements:
+1. Use Mexican Spanish vocabulary and expressions
+2. Adjust for {formality} formality level
+3. Fix any contextual errors or awkward phrasing
+4. Preserve idiomatic expressions appropriately for Mexican Spanish
+
+Respond in this format:
+TRANSLATION: [your refined translation]
+EXPLANATION: [Brief explanation of changes made and why this version fits {formality} Mexican Spanish better]"""
+    else:
+        refine_prompt = f"""You are an expert English translator. Refine the following translation and explain your changes:
+
+Original Spanish: {original_text}
+Initial English translation: {translation}
+Formality level: {formality}
+
+Requirements:
+1. Use natural English expressions
+2. Adjust for {formality} formality level
+3. Fix any contextual errors or awkward phrasing
+4. Preserve meaning while making it sound natural
+
+Respond in this format:
+TRANSLATION: [your refined translation]
+EXPLANATION: [Brief explanation of changes made and why this version fits {formality} English better]"""
+
+    try:
+        response = cerebras_client.chat_completion(
+            messages=[{"role": "user", "content": refine_prompt}],
+            max_tokens=512,
+            temperature=0.3
+        )
+        
+        # Parse response to extract translation and explanation
+        content = response.choices[0].message.content.strip()
+        
+        if "TRANSLATION:" in content and "EXPLANATION:" in content:
+            translation_part = content.split("TRANSLATION:")[1].split("EXPLANATION:")[0].strip()
+            explanation_part = content.split("EXPLANATION:")[1].strip()
+            return translation_part, explanation_part
+        else:
+            return content, "Explanation not available in expected format"
+            
+    except Exception as e:
+        return f"Refinement error: {str(e)}", ""
+
+def complete_translation(text, direction, formality):
+    if not text.strip():
+        return "", "", ""
+    
+    try:
+        # Step 1: Initial translation
+        if direction == "English to Spanish":
+            initial_translation = translate_en_to_es(text)
+            refined_translation, explanation = refine_with_llama(text, initial_translation, "en_to_es", formality)
+        else:  # Spanish to English
+            initial_translation = translate_es_to_en(text)
+            refined_translation, explanation = refine_with_llama(text, initial_translation, "es_to_en", formality)
+        
+        return initial_translation, refined_translation, explanation
+    except Exception as e:
+        return f"Error: {str(e)}", "", ""
+
+# Create Gradio interface
+with gr.Blocks(title="Mexican Spanish Translator with Explanations") as demo:
+    gr.Markdown("# Mexican Spanish Translator")
+    gr.Markdown("Powered by NLLB-200 + Llama 4 with detailed explanations for Mexican Spanish formality levels")
+    
+    with gr.Row():
+        with gr.Column(scale=2):
+            input_text = gr.Textbox(
+                label="Text to Translate",
+                placeholder="Enter text in English or Spanish...",
+                lines=6
+            )
+            
+            with gr.Row():
+                direction = gr.Dropdown(
+                    choices=["English to Spanish", "Spanish to English"],
+                    value="English to Spanish",
+                    label="Translation Direction"
+                )
+                formality = gr.Dropdown(
+                    choices=["informal", "neutral", "formal"],
+                    value="neutral",
+                    label="Formality Level (Mexican Spanish)"
+                )
+            
+            translate_btn = gr.Button("Translate", variant="primary", size="lg")
+        
+        with gr.Column(scale=2):
+            initial_output = gr.Textbox(
+                label="Initial Translation (NLLB-200)",
+                lines=2,
+                interactive=False
+            )
+            refined_output = gr.Textbox(
+                label="Refined Translation (Llama 4)",
+                lines=2,
+                interactive=False
+            )
+            explanation_output = gr.Textbox(
+                label="Explanation of Changes",
+                lines=4,
+                interactive=False
+            )
+    
+    # Connect function
+    translate_btn.click(
+        fn=complete_translation,
+        inputs=[input_text, direction, formality],
+        outputs=[initial_output, refined_output, explanation_output]
+    )
+    
+    input_text.submit(
+        fn=complete_translation,
+        inputs=[input_text, direction, formality],
+        outputs=[initial_output, refined_output, explanation_output]
+    )
+
+if __name__ == "__main__":
+    demo.launch()