Update app.py
Browse files
app.py
CHANGED
@@ -1,10 +1,21 @@
|
|
1 |
import os
|
2 |
import gradio as gr
|
|
|
3 |
import torch
|
4 |
import torch._dynamo
|
5 |
-
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|
|
6 |
from threading import Thread
|
7 |
-
import
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
8 |
|
9 |
# Desactivar TorchDynamo para evitar errores de compilación
|
10 |
torch._dynamo.config.suppress_errors = True
|
@@ -12,9 +23,19 @@ torch._dynamo.disable()
|
|
12 |
|
13 |
# Configuración
|
14 |
MODEL_ID = "somosnlp-hackathon-2025/iberotales-gemma-3-1b-it-es"
|
15 |
-
|
|
|
|
|
|
|
16 |
DEFAULT_MAX_NEW_TOKENS = 2048
|
17 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
18 |
|
19 |
# System prompt personalizado
|
20 |
DEFAULT_SYSTEM_MESSAGE = """Resuelve el siguiente problema.
|
@@ -22,16 +43,125 @@ Primero, piensa en voz alta qué debes hacer, paso por paso y de forma resumida,
|
|
22 |
Luego, da la respuesta final entre <SOLUTION> y </SOLUTION>.
|
23 |
No escribas nada fuera de ese formato."""
|
24 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
25 |
# Variables globales
|
26 |
model = None
|
27 |
tokenizer = None
|
|
|
28 |
|
29 |
def load_model():
|
30 |
"""Cargar modelo y tokenizador"""
|
31 |
global model, tokenizer
|
32 |
|
33 |
if torch.cuda.is_available():
|
34 |
-
print(f"Cargando modelo: {MODEL_ID}")
|
35 |
try:
|
36 |
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
|
37 |
model = AutoModelForCausalLM.from_pretrained(
|
@@ -40,170 +170,402 @@ def load_model():
|
|
40 |
device_map="auto",
|
41 |
trust_remote_code=True,
|
42 |
)
|
43 |
-
|
44 |
if tokenizer.pad_token is None:
|
45 |
tokenizer.pad_token = tokenizer.eos_token
|
46 |
-
|
47 |
-
print("¡Modelo cargado exitosamente!")
|
48 |
return True
|
49 |
except Exception as e:
|
50 |
-
print(f"Error
|
51 |
return False
|
52 |
else:
|
53 |
-
|
54 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
55 |
|
56 |
-
# Cargar modelo al iniciar
|
57 |
model_loaded = load_model()
|
58 |
|
59 |
-
|
60 |
-
|
61 |
-
|
62 |
-
|
63 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
64 |
max_new_tokens: int = DEFAULT_MAX_NEW_TOKENS,
|
65 |
temperature: float = 0.7,
|
66 |
top_p: float = 0.95,
|
67 |
top_k: int = 50,
|
68 |
repetition_penalty: float = 1.2,
|
69 |
-
):
|
70 |
-
"""
|
71 |
global model, tokenizer
|
72 |
|
73 |
if model is None or tokenizer is None:
|
74 |
-
|
|
|
75 |
return
|
76 |
|
77 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
78 |
|
79 |
-
|
80 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
81 |
|
82 |
-
|
83 |
-
|
84 |
-
|
|
|
85 |
|
86 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
87 |
|
|
|
|
|
|
|
|
|
88 |
try:
|
89 |
-
|
90 |
-
|
91 |
-
|
92 |
-
|
93 |
-
|
|
|
|
|
|
|
94 |
|
95 |
-
|
96 |
-
|
97 |
-
|
98 |
|
99 |
-
|
100 |
-
|
|
|
|
|
|
|
|
|
|
|
101 |
|
102 |
-
|
103 |
-
|
104 |
-
|
105 |
-
|
106 |
-
skip_special_tokens=True
|
107 |
-
)
|
108 |
|
109 |
-
|
110 |
-
|
111 |
-
|
112 |
-
|
113 |
-
"max_new_tokens": max_new_tokens,
|
114 |
-
"do_sample": True,
|
115 |
-
"top_p": top_p,
|
116 |
-
"top_k": top_k,
|
117 |
-
"temperature": temperature,
|
118 |
-
"repetition_penalty": repetition_penalty,
|
119 |
-
"pad_token_id": tokenizer.eos_token_id,
|
120 |
-
"eos_token_id": tokenizer.eos_token_id,
|
121 |
-
}
|
122 |
-
|
123 |
-
generation_thread = Thread(target=model.generate, kwargs=generate_kwargs)
|
124 |
-
generation_thread.start()
|
125 |
-
|
126 |
-
outputs = []
|
127 |
-
try:
|
128 |
-
for new_text in streamer:
|
129 |
-
outputs.append(new_text)
|
130 |
-
yield "".join(outputs)
|
131 |
-
except Exception as e:
|
132 |
-
yield f"Error durante la generación: {str(e)}"
|
133 |
-
finally:
|
134 |
-
generation_thread.join(timeout=1)
|
135 |
|
136 |
-
|
137 |
-
|
138 |
-
|
139 |
-
|
140 |
-
demo = gr.ChatInterface(
|
141 |
-
fn=generate,
|
142 |
-
title="Iberotales: Mitos y Leyendas Iberoamericanas",
|
143 |
-
description="Genera historias y personajes basados en el patrimonio cultural de Iberoamérica usando GRPO.",
|
144 |
-
chatbot=gr.Chatbot(
|
145 |
-
height=600,
|
146 |
-
show_copy_button=True,
|
147 |
-
),
|
148 |
-
textbox=gr.Textbox(
|
149 |
-
placeholder="Escribe una historia o personaje que quieras generar...",
|
150 |
-
scale=7
|
151 |
-
),
|
152 |
-
additional_inputs=[
|
153 |
-
gr.Textbox(
|
154 |
-
value=DEFAULT_SYSTEM_MESSAGE,
|
155 |
-
label="Mensaje del sistema (formato estructurado requerido)"
|
156 |
-
),
|
157 |
-
gr.Slider(
|
158 |
-
label="Máximo de tokens",
|
159 |
-
minimum=100,
|
160 |
-
maximum=MAX_MAX_NEW_TOKENS,
|
161 |
-
step=50,
|
162 |
-
value=DEFAULT_MAX_NEW_TOKENS,
|
163 |
-
),
|
164 |
-
gr.Slider(
|
165 |
-
label="Temperatura",
|
166 |
-
minimum=0.1,
|
167 |
-
maximum=2.0,
|
168 |
-
step=0.1,
|
169 |
-
value=0.7,
|
170 |
-
),
|
171 |
-
gr.Slider(
|
172 |
-
label="Top-p",
|
173 |
-
minimum=0.1,
|
174 |
-
maximum=1.0,
|
175 |
-
step=0.05,
|
176 |
-
value=0.95,
|
177 |
-
),
|
178 |
-
gr.Slider(
|
179 |
-
label="Top-k",
|
180 |
-
minimum=1,
|
181 |
-
maximum=100,
|
182 |
-
step=1,
|
183 |
-
value=50,
|
184 |
-
),
|
185 |
-
gr.Slider(
|
186 |
-
label="Penalización por repetición",
|
187 |
-
minimum=1.0,
|
188 |
-
maximum=2.0,
|
189 |
-
step=0.05,
|
190 |
-
value=1.2,
|
191 |
-
),
|
192 |
-
],
|
193 |
-
examples=[
|
194 |
-
["Crea una historia corta sobre el Pombero, un personaje de la mitología guaraní."],
|
195 |
-
["Genera un personaje basado en la leyenda del Cadejo."],
|
196 |
-
["Inventa una narrativa en torno al Nahual en un entorno contemporáneo."],
|
197 |
-
],
|
198 |
-
cache_examples=False,
|
199 |
-
)
|
200 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
201 |
if __name__ == "__main__":
|
202 |
if model_loaded:
|
203 |
-
|
204 |
-
demo.launch(
|
205 |
-
share=False,
|
206 |
-
show_error=True
|
207 |
-
)
|
208 |
else:
|
209 |
-
print("Error al cargar el modelo.
|
|
|
1 |
import os
|
2 |
import gradio as gr
|
3 |
+
from gradio import ChatMessage
|
4 |
import torch
|
5 |
import torch._dynamo
|
6 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
7 |
+
from llama_cpp import Llama
|
8 |
from threading import Thread
|
9 |
+
from huggingface_hub import hf_hub_download
|
10 |
+
import re
|
11 |
+
from typing import Iterator
|
12 |
+
|
13 |
+
# Intentar importar spaces solo si estamos en un espacio de Hugging Face
|
14 |
+
try:
|
15 |
+
import spaces
|
16 |
+
SPACES_AVAILABLE = True
|
17 |
+
except ImportError:
|
18 |
+
SPACES_AVAILABLE = False
|
19 |
|
20 |
# Desactivar TorchDynamo para evitar errores de compilación
|
21 |
torch._dynamo.config.suppress_errors = True
|
|
|
23 |
|
24 |
# Configuración
|
25 |
MODEL_ID = "somosnlp-hackathon-2025/iberotales-gemma-3-1b-it-es"
|
26 |
+
GGUF_MODEL_ID = "somosnlp-hackathon-2025/iberotales-gemma-3-1b-it-es-finetune-gguf"
|
27 |
+
GGUF_FILENAME = "gemma-3-finetune.Q8_0.gguf"
|
28 |
+
GGUF_REVISION = "main"
|
29 |
+
MAX_MAX_NEW_TOKENS = 2048
|
30 |
DEFAULT_MAX_NEW_TOKENS = 2048
|
31 |
+
|
32 |
+
# Verificar si estamos en un espacio de Hugging Face
|
33 |
+
IS_HF_SPACE = any([
|
34 |
+
os.getenv("SPACE_ID") is not None,
|
35 |
+
os.getenv("SPACE_AUTHOR_NAME") is not None,
|
36 |
+
os.getenv("SPACE_REPO_NAME") is not None,
|
37 |
+
os.getenv("SPACE_HOST") is not None,
|
38 |
+
])
|
39 |
|
40 |
# System prompt personalizado
|
41 |
DEFAULT_SYSTEM_MESSAGE = """Resuelve el siguiente problema.
|
|
|
43 |
Luego, da la respuesta final entre <SOLUTION> y </SOLUTION>.
|
44 |
No escribas nada fuera de ese formato."""
|
45 |
|
46 |
+
# Base de datos de personajes por país con banderas
|
47 |
+
PERSONAJES_POR_PAIS = {
|
48 |
+
"🇦🇷 Argentina": [
|
49 |
+
{"nombre": "La Difunta Correa", "imagen": "images/ar1.jpg", "descripcion": "Santa popular que murió de sed siguiendo a su esposo reclutado"},
|
50 |
+
{"nombre": "El Lobizón", "imagen": "images/ar2.jpg", "descripcion": "Hombre lobo de la tradición gaucha, séptimo hijo varón maldito"},
|
51 |
+
{"nombre": "La Telesita", "imagen": "images/ar3.webp", "descripcion": "Bailarina folklórica que se aparece en festivales y zambas"}
|
52 |
+
],
|
53 |
+
"🇧🇴 Bolivia": [
|
54 |
+
{"nombre": "El Tío del Cerro Rico", "imagen": "images/bo1.webp", "descripcion": "Señor de las minas que protege y castiga a los mineros"},
|
55 |
+
{"nombre": "El Ekeko", "imagen": "images/bo2.jpg", "descripcion": "Dios aymara de la abundancia y la fortuna con jorobas"},
|
56 |
+
{"nombre": "El Jichi", "imagen": "images/bo3.webp", "descripcion": "Serpiente protectora de ríos y lagunas en la cultura andina"}
|
57 |
+
],
|
58 |
+
"🇧🇷 Brasil": [
|
59 |
+
{"nombre": "Curupira", "imagen": "images/br1.jpeg", "descripcion": "Protector del bosque amazónico con pies al revés"},
|
60 |
+
{"nombre": "Saci-Pererê", "imagen": "images/br2.jpg", "descripcion": "Duende travieso de una pierna que fuma pipa"},
|
61 |
+
{"nombre": "Yebá Bëló", "imagen": "images/br3.jpg", "descripcion": "Abuela del mundo en mitología desana, creadora del universo"}
|
62 |
+
],
|
63 |
+
"🇨🇱 Chile": [
|
64 |
+
{"nombre": "Guallipén", "imagen": "images/ch1.webp", "descripcion": "Carnero gigante que habita en los ríos de Chiloé"},
|
65 |
+
{"nombre": "Colo Colo", "imagen": "images/ch2.webp", "descripcion": "Ser maléfico nacido de huevo de gallo empollado por serpiente"},
|
66 |
+
{"nombre": "Cuchivilu", "imagen": "images/ch3.webp", "descripcion": "Cerdo marino con cola de pez que habita en Chiloé"}
|
67 |
+
],
|
68 |
+
"🇨🇴 Colombia": [
|
69 |
+
{"nombre": "El Guaca", "imagen": "images/co1.jpg", "descripcion": "Espíritu guardián de tesoros enterrados por indígenas"},
|
70 |
+
{"nombre": "Chiminigagua", "imagen": "images/co2.webp", "descripcion": "Dios creador muisca, fuente de luz y vida universal"},
|
71 |
+
{"nombre": "Jukumari", "imagen": "images/co3.jpg", "descripcion": "Oso gigante mitológico de los Andes colombianos"}
|
72 |
+
],
|
73 |
+
"🇨🇷 Costa Rica": [
|
74 |
+
{"nombre": "El Micomalo", "imagen": "images/cr1.jpg", "descripcion": "Mono gigante y feroz que ataca a los cazadores"},
|
75 |
+
{"nombre": "La Carreta sin Bueyes", "imagen": "images/crr2.jpg", "descripcion": "Carreta fantasma que recorre caminos sin animales tirando"},
|
76 |
+
{"nombre": "El Sisimiqui", "imagen": "images/cr3.webp", "descripcion": "Duende pequeño y travieso de los bosques costarricenses"}
|
77 |
+
],
|
78 |
+
"🇨🇺 Cuba": [
|
79 |
+
{"nombre": "El Güije", "imagen": "images/cu1.jpg", "descripcion": "Duende negro de aguas dulces que ahoga a los bañistas"},
|
80 |
+
{"nombre": "Itiba Cahubaba", "imagen": "images/cu2.jpg", "descripcion": "Madre primordial taína de donde brotaron las aguas"},
|
81 |
+
{"nombre": "Guabancex", "imagen": "images/cu3.jpg", "descripcion": "Diosa taína de los huracanes y vientos destructores"}
|
82 |
+
],
|
83 |
+
"🇪🇨 Ecuador": [
|
84 |
+
{"nombre": "Salun", "imagen": "images/ec1.jpg", "descripcion": "Espíritu shamán shuar que guía en visiones espirituales"},
|
85 |
+
{"nombre": "Nunkui", "imagen": "images/ec2.jpg", "descripcion": "Diosa shuar de la fertilidad de la tierra y agricultura"},
|
86 |
+
{"nombre": "Yacuruna", "imagen": "images/ec3.jpg", "descripcion": "Hombre serpiente amazónico señor de las aguas dulces"}
|
87 |
+
],
|
88 |
+
"🇪🇸 España": [
|
89 |
+
{"nombre": "Ojáncanu", "imagen": "images/es1.jpg", "descripcion": "Gigante cíclope cántabro de fuerza descomunal"},
|
90 |
+
{"nombre": "Los Carantos", "imagen": "images/es2.jpg", "descripcion": "Espíritus gallegos que anuncian desgracias familiares"},
|
91 |
+
{"nombre": "Cuegle", "imagen": "images/es3.jpg", "descripcion": "Ser asturiano de tres ojos que protege casas del mal"}
|
92 |
+
],
|
93 |
+
"🇬🇹 Guatemala": [
|
94 |
+
{"nombre": "Camalotz", "imagen": "images/gu1.jpg", "descripcion": "Dios murciélago maya señor de las cuevas y la muerte"},
|
95 |
+
{"nombre": "La Siguanaba", "imagen": "images/gu2.jpg", "descripcion": "Mujer hermosa que muestra rostro de calavera a infieles"},
|
96 |
+
{"nombre": "Gucumatz", "imagen": "images/gu3.jpg", "descripcion": "Serpiente emplumada maya, equivalente a Quetzalcóatl"}
|
97 |
+
],
|
98 |
+
"🇭🇳 Honduras": [
|
99 |
+
{"nombre": "Icelaca", "imagen": "images/ho1.jpg", "descripcion": "Serpiente gigante que habita en cuevas hondureñas"},
|
100 |
+
{"nombre": "Cihuateteo", "imagen": "images/ho2.jpg", "descripcion": "Espíritus de mujeres muertas en parto, guerreras divinas"},
|
101 |
+
{"nombre": "Comelenguas", "imagen": "images/ho3.webp", "descripcion": "Demonio que devora lenguas de personas dormidas"}
|
102 |
+
],
|
103 |
+
"🇲🇽 México": [
|
104 |
+
{"nombre": "Quetzalcóatl", "imagen": "images/me1.jpg", "descripcion": "Serpiente emplumada, dios del viento y la sabiduría"},
|
105 |
+
{"nombre": "Tlacatecolotl", "imagen": "images/me2.jpg", "descripcion": "Brujo nahuatl que se transforma en búho gigante"},
|
106 |
+
{"nombre": "Tlalocan", "imagen": "images/me3.jpg", "descripcion": "Paraíso acuático de Tláloc donde van los ahogados"}
|
107 |
+
],
|
108 |
+
"🇳🇮 Nicaragua": [
|
109 |
+
{"nombre": "La Mocuana", "imagen": "images/ni1.webp", "descripcion": "Espíritu de mujer sin cabeza que vaga de noche"},
|
110 |
+
{"nombre": "El Guácimo Renco", "imagen": "images/ni2.jpg", "descripcion": "Árbol maldito que se transforma y persigue viajeros"},
|
111 |
+
{"nombre": "La Mona Bruja", "imagen": "images/ni3.jpg", "descripcion": "Bruja transformada en mona que roba niños pequeños"}
|
112 |
+
],
|
113 |
+
"🇵🇦 Panamá": [
|
114 |
+
{"nombre": "Humantahú", "imagen": "images/pa1.jpg", "descripcion": "Chamán kuna protector de la naturaleza y animales"},
|
115 |
+
{"nombre": "La Silampa", "imagen": "images/pa2.jpg", "descripcion": "Mujer fantasma que seduce y mata a hombres solitarios"},
|
116 |
+
{"nombre": "El Chivato", "imagen": "images/pa3.jpg", "descripcion": "Cabro diabólico que aparece en encrucijadas nocturnas"}
|
117 |
+
],
|
118 |
+
"🇵🇾 Paraguay": [
|
119 |
+
{"nombre": "Mbói Tata", "imagen": "images/py1.png", "descripcion": "Serpiente de fuego protectora de campos y esteros"},
|
120 |
+
{"nombre": "Urutau", "imagen": "images/py2.jpg", "descripcion": "Ave fantasma cuyo canto anuncia tragedias familiares"},
|
121 |
+
{"nombre": "Tajy", "imagen": "images/py3.jpg", "descripcion": "Espíritu del lapacho florido en la mitología guaraní"}
|
122 |
+
],
|
123 |
+
"🇵🇪 Perú": [
|
124 |
+
{"nombre": "Wiracocha", "imagen": "images/pe1.jpg", "descripcion": "Dios creador inca señor de todas las cosas vivientes"},
|
125 |
+
{"nombre": "El Muki", "imagen": "images/pe2.webp", "descripcion": "Duende minero que protege vetas de oro y plata"},
|
126 |
+
{"nombre": "Los Hermanos Ayar", "imagen": "images/pe3.jpg", "descripcion": "Cuatro hermanos fundadores míticos del imperio inca"}
|
127 |
+
],
|
128 |
+
"🇵🇹 Portugal": [
|
129 |
+
{"nombre": "El Marialva", "imagen": "images/po1.jpg", "descripcion": "Caballero galante y seductor de la tradición portuguesa"},
|
130 |
+
{"nombre": "El Pez Milagroso", "imagen": "images/po2.webp", "descripcion": "Pez mágico que concede deseos a pescadores devotos"},
|
131 |
+
{"nombre": "La Cueva da Moeda", "imagen": "images/po3.webp", "descripcion": "Cueva encantada llena de tesoros moros perdidos"}
|
132 |
+
],
|
133 |
+
"🇵🇷 Puerto Rico": [
|
134 |
+
{"nombre": "Yúcahu", "imagen": "images/pr1.jpg", "descripcion": "Dios taíno de la yuca y protector de cosechas"},
|
135 |
+
{"nombre": "Atabey", "imagen": "images/pr2.jpg", "descripcion": "Diosa madre taína de las aguas dulces y fertilidad"},
|
136 |
+
{"nombre": "Anacacuya", "imagen": "images/pr3.png", "descripcion": "Estrella taína guía de navegantes y pescadores"}
|
137 |
+
],
|
138 |
+
"🇩🇴 República Dominicana": [
|
139 |
+
{"nombre": "El Galipote", "imagen": "images/rd1.webp", "descripcion": "Brujo que se transforma en animal para hacer maldades"},
|
140 |
+
{"nombre": "El Lugaru", "imagen": "images/rd2.webp", "descripcion": "Hombre lobo dominicano que ataca ganado y personas"},
|
141 |
+
{"nombre": "Papa Legba", "imagen": "images/rd3.jpg", "descripcion": "Loa vudú guardián de encrucijadas y comunicaciones"}
|
142 |
+
],
|
143 |
+
"🇺🇾 Uruguay": [
|
144 |
+
{"nombre": "El Yasy Yateré", "imagen": "images/ur1.jpg", "descripcion": "Duende rubio protector de niños con bastón mágico"},
|
145 |
+
{"nombre": "El Ñandú Barriga Blanca", "imagen": "images/ur2.jpg", "descripcion": "Ñandú gigante y sagrado de las pampas uruguayas"},
|
146 |
+
{"nombre": "El Pombero", "imagen": "images/ur3.jpg", "descripcion": "Duende travieso protector de aves y la naturaleza"}
|
147 |
+
],
|
148 |
+
"🇻🇪 Venezuela": [
|
149 |
+
{"nombre": "Amalivaca", "imagen": "images/ve1.jpg", "descripcion": "Héroe cultural tamanaco creador del río Orinoco"},
|
150 |
+
{"nombre": "Osemma", "imagen": "images/ve2.jpg", "descripcion": "Espíritu warao de las palmeras de moriche"},
|
151 |
+
{"nombre": "La Sayona", "imagen": "images/ve3.webp", "descripcion": "Mujer vengativa que persigue y castiga a infieles"}
|
152 |
+
]
|
153 |
+
};
|
154 |
+
|
155 |
# Variables globales
|
156 |
model = None
|
157 |
tokenizer = None
|
158 |
+
current_personajes = [] # Para mantener el estado de los personajes actuales
|
159 |
|
160 |
def load_model():
|
161 |
"""Cargar modelo y tokenizador"""
|
162 |
global model, tokenizer
|
163 |
|
164 |
if torch.cuda.is_available():
|
|
|
165 |
try:
|
166 |
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
|
167 |
model = AutoModelForCausalLM.from_pretrained(
|
|
|
170 |
device_map="auto",
|
171 |
trust_remote_code=True,
|
172 |
)
|
|
|
173 |
if tokenizer.pad_token is None:
|
174 |
tokenizer.pad_token = tokenizer.eos_token
|
|
|
|
|
175 |
return True
|
176 |
except Exception as e:
|
177 |
+
print(f"Error GPU: {e}")
|
178 |
return False
|
179 |
else:
|
180 |
+
try:
|
181 |
+
local_model_path = os.path.join("models", GGUF_FILENAME)
|
182 |
+
if os.path.exists(local_model_path):
|
183 |
+
model_path = local_model_path
|
184 |
+
else:
|
185 |
+
model_path = hf_hub_download(
|
186 |
+
repo_id=GGUF_MODEL_ID,
|
187 |
+
filename=GGUF_FILENAME,
|
188 |
+
revision=GGUF_REVISION,
|
189 |
+
local_dir="./models",
|
190 |
+
force_download=False,
|
191 |
+
resume_download=True
|
192 |
+
)
|
193 |
+
tokenizer = AutoTokenizer.from_pretrained("google/gemma-3-1b-it")
|
194 |
+
model = Llama(
|
195 |
+
model_path=model_path,
|
196 |
+
n_ctx=2048,
|
197 |
+
n_threads=4,
|
198 |
+
n_gpu_layers=0
|
199 |
+
)
|
200 |
+
return True
|
201 |
+
except Exception as e:
|
202 |
+
print(f"Error GGUF: {e}")
|
203 |
+
return False
|
204 |
|
|
|
205 |
model_loaded = load_model()
|
206 |
|
207 |
+
def format_chat_history(messages: list, exclude_last_user: bool = True) -> list:
|
208 |
+
"""Formatea el historial de chat para el modelo"""
|
209 |
+
formatted_history = []
|
210 |
+
messages_to_process = messages[:]
|
211 |
+
if exclude_last_user and messages_to_process and messages_to_process[-1].get("role") == "user":
|
212 |
+
messages_to_process = messages_to_process[:-1]
|
213 |
+
|
214 |
+
for message in messages_to_process:
|
215 |
+
current_role = message.get("role")
|
216 |
+
current_content = message.get("content", "").strip()
|
217 |
+
|
218 |
+
if current_role == "assistant" and message.get("metadata"):
|
219 |
+
continue
|
220 |
+
if not current_content:
|
221 |
+
continue
|
222 |
+
|
223 |
+
if formatted_history and formatted_history[-1]["role"] == current_role:
|
224 |
+
formatted_history[-1]["content"] += "\n\n" + current_content
|
225 |
+
else:
|
226 |
+
formatted_history.append({
|
227 |
+
"role": current_role,
|
228 |
+
"content": current_content
|
229 |
+
})
|
230 |
+
|
231 |
+
return formatted_history
|
232 |
+
|
233 |
+
def stream_iberotales_response(
|
234 |
+
user_message: str,
|
235 |
+
messages: list,
|
236 |
+
system_message: str = DEFAULT_SYSTEM_MESSAGE,
|
237 |
max_new_tokens: int = DEFAULT_MAX_NEW_TOKENS,
|
238 |
temperature: float = 0.7,
|
239 |
top_p: float = 0.95,
|
240 |
top_k: int = 50,
|
241 |
repetition_penalty: float = 1.2,
|
242 |
+
) -> Iterator[list]:
|
243 |
+
"""Genera respuesta con streaming"""
|
244 |
global model, tokenizer
|
245 |
|
246 |
if model is None or tokenizer is None:
|
247 |
+
messages.append(ChatMessage(role="assistant", content="Error: Modelo no disponible."))
|
248 |
+
yield messages
|
249 |
return
|
250 |
|
251 |
+
try:
|
252 |
+
chat_history = format_chat_history(messages, exclude_last_user=True)
|
253 |
+
conversation = []
|
254 |
+
if system_message.strip():
|
255 |
+
conversation.append({"role": "system", "content": system_message.strip()})
|
256 |
+
conversation.extend(chat_history)
|
257 |
+
conversation.append({"role": "user", "content": user_message})
|
258 |
+
|
259 |
+
# Validar alternancia
|
260 |
+
for i in range(1, len(conversation)):
|
261 |
+
if conversation[i]["role"] == conversation[i-1]["role"] and conversation[i-1]["role"] != "system":
|
262 |
+
messages.append(ChatMessage(role="assistant", content="Error: Reinicia la conversación."))
|
263 |
+
yield messages
|
264 |
+
return
|
265 |
+
|
266 |
+
prompt = tokenizer.apply_chat_template(conversation, tokenize=False, add_generation_prompt=True)
|
267 |
+
response = model(
|
268 |
+
prompt,
|
269 |
+
max_tokens=max_new_tokens,
|
270 |
+
temperature=temperature,
|
271 |
+
top_p=top_p,
|
272 |
+
top_k=top_k,
|
273 |
+
repeat_penalty=repetition_penalty,
|
274 |
+
stream=True
|
275 |
+
)
|
276 |
|
277 |
+
full_response = ""
|
278 |
+
thinking_message_index = None
|
279 |
+
solution_message_index = None
|
280 |
+
in_think_block = False
|
281 |
+
in_solution_block = False
|
282 |
+
thinking_complete = False
|
283 |
+
|
284 |
+
for chunk in response:
|
285 |
+
if chunk["choices"][0]["finish_reason"] is None:
|
286 |
+
new_text = chunk["choices"][0]["text"]
|
287 |
+
full_response += new_text
|
288 |
+
|
289 |
+
# Procesar pensamiento
|
290 |
+
if "<think>" in full_response and not thinking_complete:
|
291 |
+
if not in_think_block:
|
292 |
+
in_think_block = True
|
293 |
+
if thinking_message_index is None:
|
294 |
+
messages.append(ChatMessage(
|
295 |
+
role="assistant",
|
296 |
+
content="",
|
297 |
+
metadata={"title": "🤔 Pensando..."}
|
298 |
+
))
|
299 |
+
thinking_message_index = len(messages) - 1
|
300 |
+
|
301 |
+
think_start = full_response.find("<think>") + 7
|
302 |
+
if "</think>" in full_response:
|
303 |
+
think_end = full_response.find("</think>")
|
304 |
+
current_thinking = full_response[think_start:think_end].strip()
|
305 |
+
thinking_complete = True
|
306 |
+
in_think_block = False
|
307 |
+
else:
|
308 |
+
current_thinking = full_response[think_start:].strip()
|
309 |
+
|
310 |
+
if thinking_message_index is not None:
|
311 |
+
messages[thinking_message_index] = ChatMessage(
|
312 |
+
role="assistant",
|
313 |
+
content=current_thinking,
|
314 |
+
metadata={"title": "🤔 Pensando..."}
|
315 |
+
)
|
316 |
+
yield messages
|
317 |
+
|
318 |
+
# Procesar solución
|
319 |
+
if "<SOLUTION>" in full_response:
|
320 |
+
if not in_solution_block:
|
321 |
+
in_solution_block = True
|
322 |
+
if solution_message_index is None:
|
323 |
+
messages.append(ChatMessage(role="assistant", content=""))
|
324 |
+
solution_message_index = len(messages) - 1
|
325 |
+
|
326 |
+
solution_start = full_response.find("<SOLUTION>") + 10
|
327 |
+
if "</SOLUTION>" in full_response:
|
328 |
+
solution_end = full_response.find("</SOLUTION>")
|
329 |
+
current_solution = full_response[solution_start:solution_end].strip()
|
330 |
+
in_solution_block = False
|
331 |
+
else:
|
332 |
+
current_solution = full_response[solution_start:].strip()
|
333 |
+
|
334 |
+
if solution_message_index is not None and current_solution:
|
335 |
+
messages[solution_message_index] = ChatMessage(
|
336 |
+
role="assistant",
|
337 |
+
content=current_solution
|
338 |
+
)
|
339 |
+
yield messages
|
340 |
+
|
341 |
+
# Respuesta sin formato
|
342 |
+
if full_response.strip() and solution_message_index is None:
|
343 |
+
clean_response = full_response
|
344 |
+
if "<think>" in clean_response and "</think>" in clean_response:
|
345 |
+
clean_response = re.sub(r'<think>.*?</think>', '', clean_response, flags=re.DOTALL)
|
346 |
+
if "<SOLUTION>" in clean_response and "</SOLUTION>" in clean_response:
|
347 |
+
clean_response = re.sub(r'<SOLUTION>(.*?)</SOLUTION>', r'\1', clean_response, flags=re.DOTALL)
|
348 |
+
|
349 |
+
clean_response = clean_response.strip()
|
350 |
+
if clean_response:
|
351 |
+
messages.append(ChatMessage(role="assistant", content=clean_response))
|
352 |
+
yield messages
|
353 |
+
|
354 |
+
except Exception as e:
|
355 |
+
messages.append(ChatMessage(role="assistant", content=f"Error: {str(e)}"))
|
356 |
+
yield messages
|
357 |
|
358 |
+
def user_message(msg: str, history: list) -> tuple[str, list]:
|
359 |
+
"""Añade mensaje del usuario al historial"""
|
360 |
+
history.append(ChatMessage(role="user", content=msg))
|
361 |
+
return "", history
|
362 |
|
363 |
+
def actualizar_personajes(pais_seleccionado):
|
364 |
+
"""Actualiza la galería de personajes según el país seleccionado"""
|
365 |
+
global current_personajes
|
366 |
+
personajes = PERSONAJES_POR_PAIS.get(pais_seleccionado, [])
|
367 |
+
current_personajes = personajes # Guardamos el estado actual
|
368 |
+
|
369 |
+
if not personajes:
|
370 |
+
return [], "Selecciona un país para ver sus personajes"
|
371 |
+
|
372 |
+
# Crear lista de imágenes y etiquetas para la galería
|
373 |
+
imagenes = []
|
374 |
+
for p in personajes:
|
375 |
+
if os.path.exists(p["imagen"]):
|
376 |
+
imagenes.append((p["imagen"], f"{p['nombre']}: {p['descripcion']}"))
|
377 |
+
else:
|
378 |
+
# Imagen placeholder si no existe
|
379 |
+
imagenes.append(("data:image/svg+xml;base64,PHN2ZyB3aWR0aD0iMTAwIiBoZWlnaHQ9IjEwMCIgeG1sbnM9Imh0dHA6Ly93d3cudzMub3JnLzIwMDAvc3ZnIj48cmVjdCB3aWR0aD0iMTAwIiBoZWlnaHQ9IjEwMCIgZmlsbD0iI2NjYyIvPjx0ZXh0IHg9IjUwIiB5PSI1MCIgZm9udC1mYW1pbHk9IkFyaWFsIiBmb250LXNpemU9IjEyIiBmaWxsPSIjNjY2IiB0ZXh0LWFuY2hvcj0ibWlkZGxlIiBkeT0iLjNlbSI+SW1hZ2VuPC90ZXh0Pjwvc3ZnPg==", f"{p['nombre']}: {p['descripcion']}"))
|
380 |
+
|
381 |
+
return imagenes, f"Personajes de {pais_seleccionado}"
|
382 |
|
383 |
+
def crear_prompt_desde_personaje(evt: gr.SelectData):
|
384 |
+
"""Crea un prompt basado en el personaje seleccionado"""
|
385 |
+
global current_personajes
|
386 |
+
|
387 |
try:
|
388 |
+
if evt.index is not None and evt.index < len(current_personajes):
|
389 |
+
personaje = current_personajes[evt.index]
|
390 |
+
return f"Crea una historia sobre {personaje['nombre']}, {personaje['descripcion']}" #si alguien lee esto, cambiar el dataste a cuenta en lugar de crea
|
391 |
+
else:
|
392 |
+
return "Crea una historia sobre un personaje mítico"
|
393 |
+
except Exception as e:
|
394 |
+
print(f"Error al crear prompt: {e}")
|
395 |
+
return "Crea una historia sobre un personaje mítico"
|
396 |
|
397 |
+
# Aplicar decorador @spaces.GPU si es necesario
|
398 |
+
if IS_HF_SPACE and SPACES_AVAILABLE and torch.cuda.is_available():
|
399 |
+
stream_iberotales_response = spaces.GPU(stream_iberotales_response)
|
400 |
|
401 |
+
# CSS personalizado para mejorar la apariencia
|
402 |
+
custom_css = """
|
403 |
+
.gradio-container {
|
404 |
+
max-width: 1400px !important;
|
405 |
+
margin: auto;
|
406 |
+
padding-top: 1.5rem;
|
407 |
+
}
|
408 |
|
409 |
+
#galeria .grid-wrap {
|
410 |
+
max-height: 350px;
|
411 |
+
overflow-y: auto;
|
412 |
+
}
|
|
|
|
|
413 |
|
414 |
+
#galeria .grid-container {
|
415 |
+
grid-template-columns: repeat(1, 1fr) !important;
|
416 |
+
gap: 0.5rem;
|
417 |
+
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
418 |
|
419 |
+
#galeria .thumbnail-item {
|
420 |
+
aspect-ratio: 1;
|
421 |
+
max-height: 100px;
|
422 |
+
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
423 |
|
424 |
+
#galeria .thumbnail-item img {
|
425 |
+
object-fit: cover;
|
426 |
+
width: 100%;
|
427 |
+
height: 100%;
|
428 |
+
border-radius: 8px;
|
429 |
+
}
|
430 |
+
|
431 |
+
.header-info {
|
432 |
+
background: linear-gradient(135deg, #2c3e50 0%, #1a1a2e 100%);
|
433 |
+
color: white;
|
434 |
+
padding: 1rem;
|
435 |
+
border-radius: 12px;
|
436 |
+
margin-bottom: 1rem;
|
437 |
+
text-align: center;
|
438 |
+
}
|
439 |
+
"""
|
440 |
+
|
441 |
+
# Crear la interfaz
|
442 |
+
with gr.Blocks(fill_height=True, title="Iberotales", css=custom_css) as demo:
|
443 |
+
# Header con información del proyecto
|
444 |
+
with gr.Row():
|
445 |
+
with gr.Column():
|
446 |
+
gr.HTML("""
|
447 |
+
<div class="header-info">
|
448 |
+
<h1>📚 Iberotales</h1>
|
449 |
+
<p><strong>Autor:</strong> David Quispe | <a href="https://github.com/mcdaqc/Iberotales" target="_blank" style="text-decoration: none;">GitHub</a> | <a href="https://huggingface.co/somosnlp-hackathon-2025/iberotales-gemma-3-1b-it-es" target="_blank" style="text-decoration: none;">Modelo</a> | <a href="https://huggingface.co/somosnlp-hackathon-2025/iberotales-gemma-3-1b-it-es-finetune-gguf" target="_blank" style="text-decoration: none;">GGUF</a></p>
|
450 |
+
<p><em>Alineando modelos de lenguaje con la narrativa de mitos y leyendas de Iberoamérica.</em></p>
|
451 |
+
<p><em>Hackathon SomosNLP 2025</em></p>
|
452 |
+
</div>
|
453 |
+
""")
|
454 |
+
|
455 |
+
with gr.Row():
|
456 |
+
# Panel izquierdo - Pokédex de personajes
|
457 |
+
with gr.Column(scale=1, min_width=320):
|
458 |
+
gr.Markdown("### 🗃️ Pokédex de Personajes")
|
459 |
+
|
460 |
+
pais_dropdown = gr.Dropdown(
|
461 |
+
choices=list(PERSONAJES_POR_PAIS.keys()),
|
462 |
+
value="🇵🇾 Paraguay",
|
463 |
+
label="País",
|
464 |
+
container=False
|
465 |
+
)
|
466 |
+
|
467 |
+
galeria_personajes = gr.Gallery(
|
468 |
+
value=[],
|
469 |
+
label="Personajes",
|
470 |
+
show_label=False,
|
471 |
+
elem_id="galeria",
|
472 |
+
columns=1,
|
473 |
+
rows=4,
|
474 |
+
height=350,
|
475 |
+
object_fit="cover",
|
476 |
+
preview=False # Esto evita que se expanda automáticamente
|
477 |
+
)
|
478 |
+
|
479 |
+
# Panel derecho - Chat
|
480 |
+
with gr.Column(scale=2):
|
481 |
+
chatbot = gr.Chatbot(
|
482 |
+
type="messages",
|
483 |
+
show_label=False,
|
484 |
+
height=400,
|
485 |
+
avatar_images=(None, "🏛️")
|
486 |
+
)
|
487 |
+
|
488 |
+
with gr.Row():
|
489 |
+
input_box = gr.Textbox(
|
490 |
+
placeholder="Escribe tu historia o selecciona un personaje...",
|
491 |
+
show_label=False,
|
492 |
+
scale=4,
|
493 |
+
container=False
|
494 |
+
)
|
495 |
+
send_button = gr.Button("📤", scale=1, variant="primary")
|
496 |
+
|
497 |
+
with gr.Row():
|
498 |
+
clear_button = gr.Button("🗑️ Limpiar", scale=1, size="sm")
|
499 |
+
|
500 |
+
with gr.Column(scale=3):
|
501 |
+
with gr.Row():
|
502 |
+
max_tokens = gr.Slider(100, MAX_MAX_NEW_TOKENS, DEFAULT_MAX_NEW_TOKENS, label="Tokens", container=False)
|
503 |
+
temperature = gr.Slider(0.1, 2.0, 0.7, label="Temp", container=False)
|
504 |
+
|
505 |
+
# Variables de estado
|
506 |
+
msg_store = gr.State("")
|
507 |
+
|
508 |
+
# Eventos
|
509 |
+
def submit_message(msg, history):
|
510 |
+
if not msg.strip():
|
511 |
+
return msg, history
|
512 |
+
return "", user_message(msg, history)[1]
|
513 |
+
|
514 |
+
def generate_response(msg, history, max_tok, temp):
|
515 |
+
yield from stream_iberotales_response(msg, history, DEFAULT_SYSTEM_MESSAGE, max_tok, temp)
|
516 |
+
|
517 |
+
# Actualizar personajes cuando cambia el país
|
518 |
+
pais_dropdown.change(
|
519 |
+
fn=actualizar_personajes,
|
520 |
+
inputs=[pais_dropdown],
|
521 |
+
outputs=[galeria_personajes, gr.Textbox(visible=False)]
|
522 |
+
)
|
523 |
+
|
524 |
+
# Cargar personajes iniciales
|
525 |
+
demo.load(
|
526 |
+
fn=actualizar_personajes,
|
527 |
+
inputs=[pais_dropdown],
|
528 |
+
outputs=[galeria_personajes, gr.Textbox(visible=False)]
|
529 |
+
)
|
530 |
+
|
531 |
+
# Crear prompt desde galería
|
532 |
+
galeria_personajes.select(
|
533 |
+
fn=crear_prompt_desde_personaje,
|
534 |
+
outputs=[input_box]
|
535 |
+
)
|
536 |
+
|
537 |
+
# Envío de mensajes
|
538 |
+
input_box.submit(
|
539 |
+
lambda msg, hist: (msg, submit_message(msg, hist)[1]),
|
540 |
+
inputs=[input_box, chatbot],
|
541 |
+
outputs=[msg_store, chatbot],
|
542 |
+
queue=False
|
543 |
+
).then(
|
544 |
+
generate_response,
|
545 |
+
inputs=[msg_store, chatbot, max_tokens, temperature],
|
546 |
+
outputs=chatbot
|
547 |
+
)
|
548 |
+
|
549 |
+
send_button.click(
|
550 |
+
lambda msg, hist: (msg, submit_message(msg, hist)[1]),
|
551 |
+
inputs=[input_box, chatbot],
|
552 |
+
outputs=[msg_store, chatbot],
|
553 |
+
queue=False
|
554 |
+
).then(
|
555 |
+
generate_response,
|
556 |
+
inputs=[msg_store, chatbot, max_tokens, temperature],
|
557 |
+
outputs=chatbot
|
558 |
+
)
|
559 |
+
|
560 |
+
clear_button.click(
|
561 |
+
lambda: ([], "", ""),
|
562 |
+
outputs=[chatbot, input_box, msg_store],
|
563 |
+
queue=False
|
564 |
+
)
|
565 |
+
|
566 |
+
# Lanzar aplicación
|
567 |
if __name__ == "__main__":
|
568 |
if model_loaded:
|
569 |
+
demo.launch(share=False, show_error=True)
|
|
|
|
|
|
|
|
|
570 |
else:
|
571 |
+
print("Error al cargar el modelo.")
|