t

app.py

@@ -4,7 +4,12 @@ from typo_check import css, process_input,MODEL_OPTIONS_TYPO
 from title_gen import generate_title, MODEL_OPTIONS_TITLE
 from content_gen import generate_content, MODEL_OPTIONS_CONTENT, get_default_prompt
 from instruct_dv import generate_response, MODEL_OPTIONS_INSTRUCT 
+from en_dv_latin import translate, MODEL_OPTIONS_TRANSLATE
 
+def update_textbox_direction(direction):
+    # Enable RTL only if the source language is Dhivehi (dv2*)
+    is_rtl = direction.startswith("dv2")
+    return gr.Textbox(rtl=is_rtl)
 
 # Create Gradio interface using the latest syntax
 with gr.Blocks(theme=gr.themes.Default(), css=css) as demo:
@@ -237,6 +242,49 @@ All outputs generated are synthetic, created using fine-tuned models for experim
 - The model is experimental and may not always follow instructions perfectly.
 """)
 
+        with gr.Tab("Translation Tasks"):
+            gr.Markdown("# <center>Dhivehi Translation</center>")
+            gr.Markdown("Select a translation direction and enter text to translate between Dhivehi, English and Latin script.")
+            with gr.Row():
+                instruction = gr.Dropdown(
+                    choices=["en2dv:", "dv2en:", "dv2latin:", "latin2dv:"],
+                    label="Translation Direction",
+                    value="dv2latin:"
+                )
+            with gr.Row():
+                input_text = gr.Textbox(lines=2, label="Text to Translate", rtl=True, elem_classes="textbox1")
+            with gr.Row():
+                model_choice = gr.Dropdown(choices=list(MODEL_OPTIONS_TRANSLATE.keys()), value=list(MODEL_OPTIONS_TRANSLATE.keys())[0], label="Model")
+            with gr.Row():
+                generated_response = gr.Textbox(label="Translated Text", rtl=True, elem_classes="textbox1")
+            with gr.Row():
+                max_tokens_slider = gr.Slider(10, 128, value=128, label="Max New Tokens")
+                num_beams_slider = gr.Slider(1, 10, value=4, step=10, label="Beam Size (num_beams)")
+            with gr.Row():
+                rep_penalty_slider = gr.Slider(1.0, 1.9, value=1.2, step=0.1, label="Repetition Penalty")
+                ngram_slider = gr.Slider(0, 10, value=3, step=1, label="No Repeat Ngram Size")
+            generate_btn = gr.Button("Translate")
+
+            generate_btn.click(
+                fn=translate,
+                inputs=[instruction, input_text, model_choice,max_tokens_slider, num_beams_slider, rep_penalty_slider, ngram_slider],
+                outputs=generated_response
+            )
+            gr.Examples(
+                examples=[
+                    ["dv2en:", "ދުނިޔޭގެ އެކި ކަންކޮޅުތަކުން 1.4 މިލިއަން މީހުން މައްކާއަށް ޖަމާވެފައި"],
+                    ["en2dv:", "Concerns over prepayment of GST raised in parliament"],
+                    ["dv2latin:", "ވައިބާރުވުމުން ކުޅުދުއްފުށީ އެއާޕޯޓަށް ނުޖެއްސިގެން މޯލްޑިވިއަންގެ ބޯޓެއް އެނބުރި މާލެއަށް"],
+                    ["latin2dv:", "Paakisthaanuge skoolu bahakah dhin hamalaaehgai thin kuhjakaai bodu dhe meehaku maruvehje"],
+                ],
+                inputs=[instruction, input_text],
+            )
+            gr.Markdown("""\
+**Notes:**
+- Supports translation between Dhivehi, English and Latin script
+- Model trained on news articles and common phrases
+- Translation quality may vary based on the domain of the text
+""")
 # Launch the app
 if __name__ == "__main__":
     #demo.launch(server_name="0.0.0.0", server_port=7811)

en_dv_latin.py

@@ -0,0 +1,63 @@
+import random
+import numpy as np
+import torch
+from transformers import T5Tokenizer, T5ForConditionalGeneration
+import spaces
+
+
+# Available models
+MODEL_OPTIONS_TRANSLATE = {
+    "T1DV Model": "alakxender/flan-t5-base-dhivehi-en-latin",
+}
+
+# Cache for loaded models/tokenizers
+MODEL_CACHE = {}
+
+def get_model_and_tokenizer(model_dir):
+    if model_dir not in MODEL_CACHE:
+        print(f"Loading model: {model_dir}")
+        tokenizer = T5Tokenizer.from_pretrained(model_dir)
+        model = T5ForConditionalGeneration.from_pretrained(model_dir)
+        device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+        print(f"Moving model to device: {device}")
+        model.to(device)
+        MODEL_CACHE[model_dir] = (tokenizer, model)
+    return MODEL_CACHE[model_dir]
+
+max_input_length = 128
+max_output_length = 128
+
+@spaces.GPU()
+def translate(instruction, input_text, model_choice, max_new_tokens=128, num_beams=4, repetition_penalty=1.2, no_repeat_ngram_size=3):
+    model_dir = MODEL_OPTIONS_TRANSLATE[model_choice]
+    tokenizer, model = get_model_and_tokenizer(model_dir)
+
+    combined_input = f"{instruction.strip()} {input_text.strip()}" if input_text else instruction.strip()
+    inputs = tokenizer(
+        combined_input,
+        return_tensors="pt",
+        truncation=True,
+        max_length=max_input_length
+    )
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    inputs = {k: v.to(device) for k, v in inputs.items()}
+
+    gen_kwargs = {
+        **inputs,
+        "max_length":max_new_tokens,
+        "min_length":10,
+        "num_beams":num_beams,
+        "early_stopping":True,
+        "no_repeat_ngram_size":no_repeat_ngram_size,
+        "repetition_penalty":repetition_penalty,
+        "do_sample":False,
+        "pad_token_id":tokenizer.pad_token_id,
+        "eos_token_id":tokenizer.eos_token_id
+    }
+
+
+    with torch.no_grad():
+        outputs = model.generate(**gen_kwargs)
+    decoded_output = tokenizer.decode(outputs[0], skip_special_tokens=True)
+
+    return decoded_output