Initial commit

README.md

@@ -1,8 +1,8 @@
 ---
 title: FLOR 1.3B GL
-emoji: 📊
-colorFrom: green
-colorTo: pink
+emoji: 💐
+colorFrom: blue
+colorTo: white
 sdk: gradio
 sdk_version: 4.19.2
 app_file: app.py

app.py

@@ -0,0 +1,132 @@
+import gradio as gr
+from gradio.components import Slider
+import torch
+from transformers import pipeline, AutoTokenizer, AutoModelForCausalLM
+
+# Model, information and examples ----------------------------------------------
+model_id = "proxectonos/FLOR-1.3B-GL"
+title = "Modelo de xeración de texto FLOR-1.3B-GL"
+markdown_description = """
+# FLOR-1.3B-GL
+
+🪷 **[FLOR-1.3B-GL](https://huggingface.co/proxectonos/FLOR-1.3B-GL)** is a 1.3B parameters multilingual LLM for Galician language.
+
+👀 **Learn more about FLOR-1.3B:** [HF official model card](https://huggingface.co/proxectonos/FLOR-1.3B-GL) and the [Proxecto Nós](https://nos.gal/en/proxecto-nos).
+"""
+
+short_prompts_examples = [
+    ["A receita tradicional das filloas é"], 
+    ["O neno vivía preto de"]
+]
+
+few_shot_prompts_examples = [
+    ["Responde á seguinte pregunta. \nPregunta: \"Cal é a capital de Noruega? \"\nResposta: \"A capital de Noruega é Oslo.\"\n---- \nResponde á seguinte pregunta.\nPregunta: \"Cal é a moeda de Portugal\" \nResposta: \"A moeda de Portugal é o euro.\" \n---- \nResponde á seguinte pregunta. \nPregunta: \"Cal é a capital de Suecia?\"\nResposta:"],
+    ["Extrae as entidades nomeadas do seguinte texto: \nTexto: \"Chámome Wolfgang e vivo en Berlin\" \nEntidades: Wolfgang:PER, Berlin:LOC \n ---- \nExtrae as entidades nomeadas do seguinte texto: \nTexto: \"María e Miguel non teñen ningún problema\" \nEntidades: María:PER, Miguel:PER \n---- \nExtrae as entidades nomeadas do seguinte texto: \nTexto: \"O mellor de Barcelona é o bar do meu amigo Pablo\" \nEntidades: Pablo:PER, Barcelona:LOC \n---- \nExtrae as entidades nomeadas do seguinte texto: \nTexto: \"Carlos comparte cuarto con Marc\" \nEntidades:"]
+]
+fronted_theme = 'Soft'
+
+# Model charge ---------------------------------------------------------
+model_id  = "proxectonos/FLOR-1.3B-GL"
+generator_model = pipeline("text-generation", model=model_id)
+
+# Generation functions ---------------------------------------------------------
+def remove_empty_lines(text):
+    lines = text.strip().split("\n")
+    non_empty_lines = [line for line in lines if line.strip()]
+    return "\n".join(non_empty_lines)
+
+def predict(prompt, max_length, repetition_penalty=1.3):
+    print("Dentro da xeración...")
+    prompt_length = len(generator_model.tokenizer.encode(prompt))
+    generated_text = generator_model(
+        prompt,
+        max_length=prompt_length + max_length, 
+        pad_token_id=generator_model.tokenizer.eos_token_id, 
+        repetition_penalty=repetition_penalty)
+
+    generated_sequence = generated_text[0]['generated_text']
+    if generated_sequence is  None:
+        gr.Warning('Inference endpoint is not available right now. Please try again later.')
+        return
+    
+    generated_sequence = remove_empty_lines(generated_sequence)
+    print("Xeración completada")
+    return generated_sequence
+
+# Gradio app ---------------------------------------------------------
+def clear(): 
+    return (
+        None, 
+        None,
+        gr.update(value=20),
+        gr.update(value=1.3)
+    )
+def pass_to_input(generated_gl):
+    return (
+        gr.update(value=generated_gl),
+        None,
+    )
+
+def gradio_app():
+    with gr.Blocks(theme=fronted_theme) as demo:
+        with gr.Row():
+            with gr.Column(scale=0.1):
+                gr.HTML('<img src="https://huggingface.co/spaces/proxectonos/README/resolve/main/title-card.png" width="100%" style="border-radius: 0.75rem;">')
+            with gr.Column():
+                gr.Markdown(markdown_description)
+
+        with gr.Row(equal_height=True):
+            with gr.Column():
+                text_gl = gr.Textbox(label="Input", 
+                                     lines=6, placeholder="e.g. O neno vai a escola con ")
+                with gr.Row(variant="panel"):
+                        with gr.Accordion("Model parameters", open=False):
+                            max_length = Slider(
+                                minimum=1,
+                                maximum=200,
+                                step=1,
+                                value=30,
+                                label="Max tokens"
+                            )
+                            repetition_penalty = Slider(
+                                minimum=0.1,
+                                maximum=4,
+                                step=0.1,
+                                value=1.3,
+                                label="Repetition penalty"
+                            )
+                generator_btn = gr.Button(value="Generate",variant='primary')
+            with gr.Column():
+                generated_gl = gr.Textbox(label="Output", 
+                                          lines=6, 
+                                          placeholder="Generated text will appear here",
+                                          interactive=False,
+                                          show_copy_button=True)
+                pass_btn = gr.Button(value="Pass text to input")
+                clean_btn = gr.Button(value="Clean")
+
+        generator_btn.click(predict, inputs=[text_gl,max_length, repetition_penalty], outputs=generated_gl, api_name="generate-flor-gl")
+        clean_btn.click(fn=clear, inputs=[], outputs=[text_gl, generated_gl, max_length, repetition_penalty], queue=False, api_name=False)
+        pass_btn.click(fn=pass_to_input, inputs=[generated_gl], outputs=[text_gl,generated_gl], queue=False, api_name=False)
+        
+        with gr.Row():
+            with gr.Column(scale=0.5):
+                gr.Examples(
+                    label = "Short prompts",
+                    examples = short_prompts_examples,
+                    inputs = [text_gl,max_length, repetition_penalty],
+                    outputs = generated_gl,
+                    fn =predict
+                )
+                gr.Examples(
+                    label = "Few-shot prompts",
+                    examples = few_shot_prompts_examples,
+                    inputs = [text_gl,max_length, repetition_penalty],
+                    outputs = generated_gl,
+                    fn =predict
+                )
+
+    demo.launch()
+
+if __name__ == "__main__":
+    gradio_app()

requirements.txt

@@ -0,0 +1,3 @@
+transformers
+torch
+accelerate