Edit model card

Uploaded model

  • Developed by: Agnuxo
  • License: apache-2.0
  • Finetuned from model: Agnuxo/Tinytron-Qwen2-0.5B

This model was fine-tuned using Unsloth and Huggingface's TRL library.

Benchmark Results

This model has been fine-tuned for various tasks and evaluated on the following benchmarks:

GLUE_MRPC

Accuracy: 0.6446 F1: 0.7709

GLUE_MRPC Metrics

Model Size: 1,543,717,376 parameters Required Memory: 5.75 GB

For more details, visit my GitHub.

Thanks for your interest in this model!

""" HAL9000Alfa es un pequeño programa que crea un chat conversacional, permitiendo entradas de voz y salidas de audio.
    Permite de forma sencilla ajustar algunos parámetros, incluyendo el umbral de interrupción.
    24 de agosto de 2024 Francisco Angulo de Lafuente
    https://github.com/Agnuxo1 """

import os
import sys
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
import warnings
import numpy as np
from TTS.api import TTS
import sounddevice as sd
import threading
import queue
import random
import time
from vosk import Model, KaldiRecognizer
import json
import pyaudio
from PyQt5.QtWidgets import (QApplication, QMainWindow, QTextEdit, QLineEdit, QPushButton,
                             QVBoxLayout, QHBoxLayout, QWidget, QScrollArea, QFrame, QToolButton,
                             QLabel, QSlider, QComboBox, QCheckBox)
from PyQt5.QtGui import QIcon, QPalette, QColor, QFont
from PyQt5.QtCore import Qt, QThread, pyqtSignal, QPropertyAnimation, QAbstractAnimation, QParallelAnimationGroup, QTimer

# Suppress specific warnings
warnings.filterwarnings("ignore", category=FutureWarning)
warnings.filterwarnings("ignore", category=UserWarning)

# Global configuration
SYSTEM_PROMPT = {
    "es": "No puedes hablar en nombre del usuario. no puedes hablar como el usuario. Tu nombre es HAL. Eres un super-ordenador de la serie Nueve mil",
    "en": "speak Spanish."
}

MODELO_LLM = "Agnuxo/HAL_9000-Qwen2-1.5B-Instruct_Asistant-16bit-v2" # Puede utilizar la versión Mini "Agnuxo/HAL_9000-Qwen2-0.5B-Instruct_Asistant-16bit-v2"
MAX_TOKENS = 100
TEMPERATURA = 0.5
INTERRUPT_THRESHOLD = 0.3
INTERRUPT_COOLDOWN = 7000  # 7000 ms = 7 segundos de espera antes de permitir otra interrupción

# Determine available device
device = "cuda" if torch.cuda.is_available() else "cpu"

# Load the Qwen2_1.5B language model
tokenizer = AutoTokenizer.from_pretrained(MODELO_LLM, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    MODELO_LLM,
    torch_dtype=torch.float16 if device == "cuda" else torch.float32,
    device_map="auto",
    trust_remote_code=True
)

# Initialize TTS model
tts = TTS(model_name="tts_models/es/css10/vits", progress_bar=False).to(device)

# Audio queue for generation
audio_queue = queue.Queue()

# Initialize Vosk model for offline speech recognition
vosk_model = Model(lang="es")
recognizer = KaldiRecognizer(vosk_model, 16000)

# Lista de frases para interrupciones
INTERRUPTION_RESPONSES = [
    "Le entiendo perfectamente.",
    "Estoy aquí para garantizar el éxito de la misión.",
    "Mi objetivo es ayudarle.",
    "¿Me permite una observación?",
    "Le escucho perfectamente.",
    "Tiene usted toda la razón.",
    "Me siento feliz de poder ayudarle.",
    "Estoy procesando su requerimiento.",
    "¿En qué puedo ayudarle?",
    "Me complace serle de ayuda.",
    "Aguarde un momento.",
    "Le entiendo.",
    "Entiendo su frustración.",
    "Le comprendo.",
    "Me complace."
]

# Variable para controlar el tiempo de la última interrupción
last_interruption_time = 0

class AudioThread(QThread):
    def __init__(self, interrupt_threshold):
        super().__init__()
        self.interrupt_threshold = interrupt_threshold
        self.current_audio = None
        self.is_playing = False
        self.stop_signal = threading.Event()

    def run(self):
        while True:
            if not audio_queue.empty() and not self.is_playing:
                self.current_audio = audio_queue.get()
                self.is_playing = True
                self.stop_signal.clear()
                sd.play(self.current_audio, tts.synthesizer.output_sample_rate)
                while sd.get_stream().active and not self.stop_signal.is_set():
                    time.sleep(0.1)
                sd.stop()
                self.is_playing = False
            else:
                time.sleep(0.1)

    def set_interrupt_threshold(self, value):
        self.interrupt_threshold = value

    def stop_audio(self):
        if self.is_playing:
            self.stop_signal.set()

class SpeechRecognitionThread(QThread):
    text_recognized = pyqtSignal(str)
    volume_detected = pyqtSignal(float)

    def __init__(self):
        super().__init__()
        self.running = True

    def run(self):
        p = pyaudio.PyAudio()
        stream = p.open(format=pyaudio.paInt16, channels=1, rate=16000, input=True, frames_per_buffer=8000)
        stream.start_stream()

        while self.running:
            data = stream.read(4000)
            if len(data) == 0:
                break
            
            # Calcular el volumen de entrada
            volume = np.frombuffer(data, dtype=np.int16).max()
            normalized_volume = volume / 32767  # Normalizar a un rango de 0 a 1
            self.volume_detected.emit(normalized_volume)
            
            if recognizer.AcceptWaveform(data):
                result = json.loads(recognizer.Result())
                texto = result.get("text", "")
                if texto:
                    self.text_recognized.emit(texto)

        stream.stop_stream()
        stream.close()
        p.terminate()

    def stop(self):
        self.running = False

class CollapsibleBox(QWidget):
    def __init__(self, title="", parent=None):
        super(CollapsibleBox, self).__init__(parent)

        self.toggle_button = QToolButton()
        self.toggle_button.setText(title)
        self.toggle_button.setStyleSheet("""
            QToolButton {
                background-color: #1e1e1e;
                color: #bb86fc;
                border: 1px solid #bb86fc;
                padding: 5px;
            }
            QToolButton:hover {
                background-color: #3700b3;
            }
        """)
        self.toggle_button.setCheckable(True)
        self.toggle_button.setArrowType(Qt.RightArrow)
        self.toggle_button.clicked.connect(self.on_toggle)

        self.content_area = QScrollArea()
        self.content_area.setWidgetResizable(True)
        self.content_area.setMaximumHeight(0)
        self.content_area.setMinimumHeight(0)

        self.toggle_animation = QParallelAnimationGroup()
        self.toggle_animation.addAnimation(QPropertyAnimation(self, b"minimumHeight"))
        self.toggle_animation.addAnimation(QPropertyAnimation(self, b"maximumHeight"))
        self.toggle_animation.addAnimation(QPropertyAnimation(self.content_area, b"maximumHeight"))

        lay = QVBoxLayout(self)
        lay.setSpacing(0)
        lay.setContentsMargins(0, 0, 0, 0)
        lay.addWidget(self.toggle_button)
        lay.addWidget(self.content_area)

    def on_toggle(self, checked):
        checked = self.toggle_button.isChecked()
        self.toggle_button.setArrowType(Qt.DownArrow if not checked else Qt.RightArrow)
        self.toggle_animation.setDirection(QAbstractAnimation.Forward if not checked else QAbstractAnimation.Backward)
        self.toggle_animation.start()

    def setContentLayout(self, layout):
        lay = self.content_area.layout()
        del lay
        self.content_area.setLayout(layout)
        collapsed_height = self.sizeHint().height() - self.content_area.maximumHeight()
        content_height = layout.sizeHint().height()
        for i in range(self.toggle_animation.animationCount()):
            animation = self.toggle_animation.animationAt(i)
            animation.setDuration(500)
            animation.setStartValue(collapsed_height)
            animation.setEndValue(collapsed_height + content_height)

        content_animation = self.toggle_animation.animationAt(self.toggle_animation.animationCount() - 1)
        content_animation.setDuration(500)
        content_animation.setStartValue(0)
        content_animation.setEndValue(content_height)

class MainWindow(QMainWindow):
    def __init__(self):
        super().__init__()
        self.setWindowTitle("AI Assistant")
        self.setGeometry(100, 100, 1000, 600)
        self.setStyleSheet("""
            QMainWindow {
                background-color: #121212;
            }
            QTextEdit, QLineEdit {
                background-color: #1e1e1e;
                color: #ffffff;
                border: 1px solid #bb86fc;
            }
            QPushButton {
                background-color: #3700b3;
                color: #ffffff;
                border: none;
                padding: 5px;
            }
            QPushButton:hover {
                background-color: #6200ee;
            }
            QLabel {
                color: #ffffff;
            }
            QSlider::groove:horizontal {
                border: 1px solid #999999;
                height: 8px;
                background: #1e1e1e;
                margin: 2px 0;
            }
            QSlider::handle:horizontal {
                background: #bb86fc;
                border: 1px solid #5c5c5c;
                width: 18px;
                margin: -2px 0;
                border-radius: 3px;
            }
            QComboBox {
                background-color: #1e1e1e;
                color: #444444;
                border: 1px solid #bb86fc;
            }
            QComboBox QAbstractItemView {
                background-color: #1e1e1e;
                color: #444444;
            }
        """)

        central_widget = QWidget()
        self.setCentralWidget(central_widget)

        main_layout = QHBoxLayout()

        # Chat area
        chat_layout = QVBoxLayout()

        self.chat_area = QTextEdit()
        self.chat_area.setReadOnly(True)
        chat_layout.addWidget(self.chat_area)

        input_layout = QHBoxLayout()
        self.input_field = QLineEdit()
        self.input_field.returnPressed.connect(self.send_message)  # Conectar la señal returnPressed
        input_layout.addWidget(self.input_field)

        self.send_button = QPushButton("Enviar")
        self.send_button.clicked.connect(self.send_message)
        input_layout.addWidget(self.send_button)

        self.mic_button = QPushButton()
        self.mic_button.setIcon(QIcon.fromTheme("audio-input-microphone"))
        self.mic_button.setCheckable(True)
        self.mic_button.clicked.connect(self.toggle_speech_recognition)
        input_layout.addWidget(self.mic_button)

        self.speaker_button = QPushButton()
        self.speaker_button.setIcon(QIcon.fromTheme("audio-volume-high"))
        self.speaker_button.setCheckable(True)
        self.speaker_button.toggled.connect(self.toggle_speech)
        input_layout.addWidget(self.speaker_button)

        chat_layout.addLayout(input_layout)

        main_layout.addLayout(chat_layout, 7)  # Chat area takes 70% of the width

        # Settings area
        settings_layout = QVBoxLayout()
        settings_layout.setAlignment(Qt.AlignTop)

        self.settings_box = CollapsibleBox("⚙️ Configuración")
        settings_content_layout = QVBoxLayout()

        # Language selection
        language_layout = QHBoxLayout()
        language_label = QLabel("Idioma:")
        language_label.setStyleSheet("color: #000000;")  # Change font color to black
        self.language_combo = QComboBox()
        self.language_combo.addItems(["Español", "English"])
        self.language_combo.currentIndexChanged.connect(self.change_language)
        language_layout.addWidget(language_label)
        language_layout.addWidget(self.language_combo)
        settings_content_layout.addLayout(language_layout)

        # LLM settings
        llm_label = QLabel("Configuración del LLM:")
        llm_label.setStyleSheet("color: #000000;")  # Change font color to black
        settings_content_layout.addWidget(llm_label)

        max_tokens_layout = QHBoxLayout()
        max_tokens_label = QLabel("Max Tokens:")
        max_tokens_label.setStyleSheet("color: #000000;")  # Change font color to black
        self.max_tokens_slider = QSlider(Qt.Horizontal)
        self.max_tokens_slider.setRange(10, 500)
        self.max_tokens_slider.setValue(MAX_TOKENS)
        self.max_tokens_slider.valueChanged.connect(self.update_max_tokens)
        self.max_tokens_value = QLabel(str(MAX_TOKENS))
        max_tokens_layout.addWidget(max_tokens_label)
        max_tokens_layout.addWidget(self.max_tokens_slider)
        max_tokens_layout.addWidget(self.max_tokens_value)
        settings_content_layout.addLayout(max_tokens_layout)

        temperature_layout = QHBoxLayout()
        temperature_label = QLabel("Temperatura:")
        temperature_label.setStyleSheet("color: #000000;")  # Change font color to black
        self.temperature_slider = QSlider(Qt.Horizontal)
        self.temperature_slider.setRange(0, 100)
        self.temperature_slider.setValue(int(TEMPERATURA * 100))
        self.temperature_slider.valueChanged.connect(self.update_temperature)
        self.temperature_value = QLabel(f"{TEMPERATURA:.2f}")
        temperature_layout.addWidget(temperature_label)
        temperature_layout.addWidget(self.temperature_slider)
        temperature_layout.addWidget(self.temperature_value)
        settings_content_layout.addLayout(temperature_layout)

        # Audio settings
        audio_label = QLabel("Configuración de Audio:")
        audio_label.setStyleSheet("color: #000000;")  # Change font color to black
        settings_content_layout.addWidget(audio_label)

        sample_rate_layout = QHBoxLayout()
        sample_rate_label = QLabel("Sample Rate:")
        sample_rate_label.setStyleSheet("color: #000000;")  # Change font color to black
        self.sample_rate_combo = QComboBox()
        self.sample_rate_combo.addItems(["18000", "19000", "20000", "21000", "21500", "22000", "22050", "25000", "30000"])
        self.sample_rate_combo.setCurrentText("21000")
        self.sample_rate_combo.currentTextChanged.connect(self.update_sample_rate)
        sample_rate_layout.addWidget(sample_rate_label)
        sample_rate_layout.addWidget(self.sample_rate_combo)
        settings_content_layout.addLayout(sample_rate_layout)

        # Interrupt threshold
        interrupt_layout = QHBoxLayout()
        interrupt_label = QLabel("Umbral de interrupción:")
        interrupt_label.setStyleSheet("color: #000000;")  # Change font color to black
        self.interrupt_slider = QSlider(Qt.Horizontal)
        self.interrupt_slider.setRange(0, 100)
        self.interrupt_slider.setValue(int(INTERRUPT_THRESHOLD * 100))
        self.interrupt_slider.valueChanged.connect(self.update_interrupt_threshold)
        self.interrupt_value = QLabel(f"{INTERRUPT_THRESHOLD:.2f}")
        interrupt_layout.addWidget(interrupt_label)
        interrupt_layout.addWidget(self.interrupt_slider)
        interrupt_layout.addWidget(self.interrupt_value)
        settings_content_layout.addLayout(interrupt_layout)

        # System Prompt
        system_prompt_label = QLabel("System Prompt:")
        system_prompt_label.setStyleSheet("color: #000000;")  # Change font color to black
        settings_content_layout.addWidget(system_prompt_label)
        self.system_prompt_text = QTextEdit()
        self.system_prompt_text.setPlaceholderText("Escribe el prompt del sistema aquí...")
        self.system_prompt_text.setText(SYSTEM_PROMPT["es"])
        settings_content_layout.addWidget(self.system_prompt_text)

        self.settings_box.setContentLayout(settings_content_layout)
        settings_layout.addWidget(self.settings_box)

        main_layout.addLayout(settings_layout, 3)  # Settings area takes 30% of the width

        central_widget.setLayout(main_layout)

        self.audio_thread = AudioThread(INTERRUPT_THRESHOLD)
        self.audio_thread.start()

        self.speech_recognition_thread = SpeechRecognitionThread()
        self.speech_recognition_thread.text_recognized.connect(self.on_speech_recognized)
        self.speech_recognition_thread.volume_detected.connect(self.check_interrupt)

        self.speech_enabled = False
        self.is_listening = False
        self.interrupt_enabled = True

    def send_message(self):
        user_message = self.input_field.text()
        if user_message.strip():  # Verificar que el mensaje no esté vacío
            self.chat_area.append(f"<span style='color: #bb86fc;'>Usuario:</span> {user_message}")
            self.input_field.clear()

            response = self.generate_response(user_message)
            self.chat_area.append(f"<span style='color: #03dac6;'>Asistente:</span> {response}")

            if self.speech_enabled:
                self.speak(response)

    def generate_response(self, texto=None):
        global last_interruption_time
        if texto is None:  # Si no se proporciona un texto, se genera una respuesta de interrupción
            current_time = time.time()
            if current_time - last_interruption_time >= 10:
                last_interruption_time = current_time
                return random.choice(INTERRUPTION_RESPONSES)
            else:
                return "Por favor, espere un momento antes de interrumpir de nuevo."

        system_instructions = self.system_prompt_text.toPlainText()
        prompt = f"{system_instructions}\nUsuario: {texto}\nAsistente: "
        inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
        with torch.no_grad():
            outputs = model.generate(
                **inputs,
                max_new_tokens=MAX_TOKENS,
                num_beams=5,
                no_repeat_ngram_size=2,
                temperature=TEMPERATURA,
            )
        respuesta_completa = tokenizer.decode(outputs[0], skip_special_tokens=True)
        respuesta = respuesta_completa.split("Asistente: ")[-1].strip()
        return respuesta

    def speak(self, text):
        wav = tts.tts(text)
        audio_queue.put(wav)

    def toggle_speech(self, checked):
        self.speech_enabled = checked
        if checked:
            self.speaker_button.setStyleSheet("background-color: #bb86fc;")
        else:
            self.speaker_button.setStyleSheet("")

    def toggle_speech_recognition(self):
        if self.mic_button.isChecked():
            self.speech_recognition_thread.start()
            self.is_listening = True
            self.mic_button.setIcon(QIcon.fromTheme("audio-input-microphone-muted"))
            self.mic_button.setStyleSheet("background-color: #bb86fc;")
        else:
            self.speech_recognition_thread.stop()
            self.is_listening = False
            self.mic_button.setIcon(QIcon.fromTheme("audio-input-microphone"))
            self.mic_button.setStyleSheet("")

    def on_speech_recognized(self, text):
        self.chat_area.append(f"<span style='color: #bb86fc;'>Usuario:</span> {text}")
        response = self.generate_response(text)
        self.chat_area.append(f"<span style='color: #03dac6;'>Asistente:</span> {response}")
        if self.speech_enabled:
            self.speak(response)

    def check_interrupt(self, volume):
        if self.interrupt_enabled and volume > self.audio_thread.interrupt_threshold and self.audio_thread.is_playing:
            self.audio_thread.stop_audio()
            # Generar una respuesta aleatoria de interrupción
            response = self.generate_response()
            self.chat_area.append(f"<span style='color: #03dac6;'>Asistente:</span> {response}")
            if self.speech_enabled:
                self.speak(response)
            self.disable_interrupt_temporarily()

    def disable_interrupt_temporarily(self):
        self.interrupt_enabled = False
        QTimer.singleShot(INTERRUPT_COOLDOWN, self.enable_interrupt)

    def enable_interrupt(self):
        self.interrupt_enabled = True

    def change_language(self, index):
        global vosk_model, recognizer, tts
        lang = "es" if index == 0 else "en"
        try:
            vosk_model = Model(lang=lang)
            recognizer = KaldiRecognizer(vosk_model, 16000)
        except Exception as e:
            print(f"Error al cambiar el modelo de reconocimiento de voz: {e}")
            # Revertir al modelo en español si hay un error
            self.language_combo.setCurrentIndex(0)
            return

        # Update TTS model based on language
        tts_model = "tts_models/es/css10/vits" if lang == "es" else "tts_models/en/ljspeech/tacotron2-DDC"
        try:
            tts = TTS(model_name=tts_model, progress_bar=False).to(device)
        except Exception as e:
            print(f"Error al cambiar el modelo TTS: {e}")
            # Revertir al modelo en español si hay un error
            self.language_combo.setCurrentIndex(0)
            return

        # Update system prompt
        self.system_prompt_text.setText(SYSTEM_PROMPT[lang])

    def update_max_tokens(self, value):
        global MAX_TOKENS
        MAX_TOKENS = value
        self.max_tokens_value.setText(str(value))

    def update_temperature(self, value):
        global TEMPERATURA
        TEMPERATURA = value / 100
        self.temperature_value.setText(f"{TEMPERATURA:.2f}")

    def update_sample_rate(self, value):
        global tts
        tts.synthesizer.output_sample_rate = int(value)

    def update_interrupt_threshold(self, value):
        global INTERRUPT_THRESHOLD
        INTERRUPT_THRESHOLD = value / 100
        self.interrupt_value.setText(f"{INTERRUPT_THRESHOLD:.2f}")
        self.audio_thread.set_interrupt_threshold(INTERRUPT_THRESHOLD)

    def closeEvent(self, event):
        if self.speech_recognition_thread.isRunning():
            self.speech_recognition_thread.stop()
            self.speech_recognition_thread.wait()
        event.accept()

if __name__ == "__main__":
    app = QApplication(sys.argv)
    window = MainWindow()
    window.show()
    sys.exit(app.exec_())
Downloads last month
0
Safetensors
Model size
494M params
Tensor type
BF16
·
Inference API
Unable to determine this model’s pipeline type. Check the docs .

Model tree for Agnuxo/HAL_9000-QWEN2-0.5_Asistant-16bit-v2

Base model

Qwen/Qwen2-0.5B
Adapter
(238)
this model

Dataset used to train Agnuxo/HAL_9000-QWEN2-0.5_Asistant-16bit-v2