Update app.py

app.py

@@ -4,6 +4,7 @@ import torch
 import json
 import time
 import gc
+import re
 from fastapi import FastAPI, Request
 from fastapi.responses import HTMLResponse, StreamingResponse
 from fastapi.middleware.cors import CORSMiddleware
@@ -143,27 +144,27 @@ class MTPModel(nn.Module):
         logits = self.lm_head(x)
         return logits
     
-    def generate(self, input_ids, max_new_tokens=100, temperature=0.8, top_k=50, top_p=0.9, repetition_penalty=1.1):
-        """Método de generación compatible con la interfaz"""
+    def generate(self, input_ids, max_new_tokens=100, temperature=0.7, top_k=50, top_p=0.9, repetition_penalty=1.1, eos_token_id=3):
+        """Método de generación mejorado con parada limpia"""
         generated = input_ids
+        eos_detected = False
         
         for _ in range(max_new_tokens):
-            # Obtener logits para el último token
             with torch.no_grad():
                 logits = self(generated)
                 next_logits = logits[0, -1, :] / temperature
             
-            # Aplicar repetition penalty
+            # Repetition penalty
             if repetition_penalty != 1.0:
                 for token_id in set(generated[0].tolist()):
                     next_logits[token_id] /= repetition_penalty
             
-            # Top-k filtering
+            # Top-k
             if top_k > 0:
                 indices_to_remove = next_logits < torch.topk(next_logits, top_k)[0][..., -1, None]
                 next_logits[indices_to_remove] = float('-inf')
             
-            # Top-p filtering
+            # Top-p
             if top_p < 1.0:
                 sorted_logits, sorted_indices = torch.sort(next_logits, descending=True)
                 cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
@@ -173,17 +174,23 @@ class MTPModel(nn.Module):
                 indices_to_remove = sorted_indices[sorted_indices_to_remove]
                 next_logits[indices_to_remove] = float('-inf')
             
-            # Sampling
             probs = F.softmax(next_logits, dim=-1)
             next_token = torch.multinomial(probs, num_samples=1).item()
             
-            # Parar en EOS
-            if next_token == 3:  # EOS ID para SentencePiece
+            # Detener en EOS o tokens sospechosos
+            if next_token == eos_token_id:
+                eos_detected = True
                 break
             
+            # Detener si detectamos repetición excesiva del mismo token
+            if len(generated[0]) > 10:
+                last_tokens = generated[0][-10:].tolist()
+                if len(set(last_tokens)) == 1:
+                    break
+            
             generated = torch.cat([generated, torch.tensor([[next_token]], device=generated.device)], dim=1)
         
-        return generated
+        return generated, eos_detected
 
 # ======================
 # DESCARGA Y CARGA DEL MODELO
@@ -216,12 +223,16 @@ tokenizer_path = os.path.join(repo_path, "mtp_tokenizer.model")
 sp = spm.SentencePieceProcessor()
 sp.load(tokenizer_path)
 VOCAB_SIZE = sp.get_piece_size()
+EOS_TOKEN_ID = sp.eos_id()
+BOS_TOKEN_ID = sp.bos_id()
 
 # Actualizar vocab_size en config
 config["vocab_size"] = VOCAB_SIZE
 
 print(f"🧠 Inicializando modelo MTP-1.1...")
 print(f"   → Vocabulario: {VOCAB_SIZE}")
+print(f"   → EOS token ID: {EOS_TOKEN_ID}")
+print(f"   → BOS token ID: {BOS_TOKEN_ID}")
 print(f"   → Dimensión: {config['d_model']}")
 print(f"   → Capas: {config['n_layers']}")
 print(f"   → Heads: {config['n_heads']}")
@@ -252,6 +263,69 @@ if DEVICE == "cpu":
 param_count = sum(p.numel() for p in model.parameters())
 print(f"✅ Modelo cargado: {param_count:,} parámetros ({param_count/1e6:.1f}M)")
 
+# ======================
+# FUNCIONES DE LIMPIEZA DE RESPUESTAS
+# ======================
+def clean_response(text: str, original_prompt: str = None) -> str:
+    """Limpia la respuesta generada eliminando basura y repeticiones"""
+    if not text:
+        return "Lo siento, no pude generar una respuesta."
+    
+    # Eliminar el prompt original si aparece al inicio
+    if original_prompt:
+        prompt_clean = original_prompt.strip().lower()
+        text_lower = text.lower()
+        if text_lower.startswith(prompt_clean):
+            text = text[len(original_prompt):].strip()
+        elif prompt_clean in text_lower[:50]:
+            # Buscar después del prompt
+            idx = text_lower.find(prompt_clean)
+            if idx != -1:
+                text = text[idx + len(original_prompt):].strip()
+    
+    # Eliminar partes que contienen "###"
+    if "###" in text:
+        text = text.split("###")[0].strip()
+    
+    # Eliminar repeticiones absurdas (patrones como "xxx" repetido)
+    words = text.split()
+    if len(words) > 10:
+        unique_words = []
+        last_word = None
+        repeat_count = 0
+        for w in words:
+            if w == last_word:
+                repeat_count += 1
+                if repeat_count > 2:
+                    continue
+            else:
+                repeat_count = 0
+            unique_words.append(w)
+            last_word = w
+        text = " ".join(unique_words)
+    
+    # Eliminar fragmentos que parecen basura (patrones sin sentido)
+    garbage_patterns = [
+        r'[a-z]{20,}',  # Palabras muy largas sin sentido
+        r'\d{5,}',       # Números muy largos
+        r'[^\w\s\.\,\!\?\-áéíóúüñ]{10,}',  # Caracteres extraños repetidos
+    ]
+    for pattern in garbage_patterns:
+        text = re.sub(pattern, '', text)
+    
+    # Limpiar espacios múltiples
+    text = re.sub(r'\s+', ' ', text).strip()
+    
+    # Capitalizar primera letra
+    if text and len(text) > 0:
+        text = text[0].upper() + text[1:] if len(text) > 1 else text.upper()
+    
+    # Si la respuesta es demasiado corta o vacía, dar mensaje por defecto
+    if len(text) < 3:
+        return "Entendido. ¿Algo más en lo que pueda ayudarte?"
+    
+    return text
+
 # ======================
 # API CONFIG
 # ======================
@@ -270,8 +344,8 @@ app.add_middleware(
 
 class PromptRequest(BaseModel):
     text: str = Field(..., max_length=2000, description="Texto de entrada")
-    max_tokens: int = Field(default=150, ge=10, le=300, description="Tokens máximos a generar")
-    temperature: float = Field(default=0.7, ge=0.1, le=2.0, description="Temperatura de muestreo")
+    max_tokens: int = Field(default=100, ge=10, le=200, description="Tokens máximos a generar")
+    temperature: float = Field(default=0.7, ge=0.1, le=1.5, description="Temperatura de muestreo")
     top_k: int = Field(default=50, ge=1, le=100, description="Top-k sampling")
     top_p: float = Field(default=0.9, ge=0.1, le=1.0, description="Top-p (nucleus) sampling")
     repetition_penalty: float = Field(default=1.1, ge=1.0, le=2.0, description="Penalización por repetición")
@@ -310,50 +384,78 @@ async def generate(req: PromptRequest):
     global ACTIVE_REQUESTS
     ACTIVE_REQUESTS += 1
     
-    dyn_max_tokens = req.max_tokens
-    dyn_temperature = req.temperature
-    
-    if ACTIVE_REQUESTS > 2:
-        print(f"⚠️ Carga alta ({ACTIVE_REQUESTS} requests). Ajustando parámetros.")
-        dyn_max_tokens = min(dyn_max_tokens, 120)
-        dyn_temperature = max(0.5, dyn_temperature * 0.9)
-
-    user_input = req.text.strip()
-    if not user_input:
-        ACTIVE_REQUESTS -= 1
-        return {"reply": "", "tokens_generated": 0}
-
-    full_prompt = build_prompt(user_input)
-    tokens = [tokenizer_wrapper.bos_id()] + tokenizer_wrapper.encode(full_prompt)
-    input_ids = torch.tensor([tokens], device=DEVICE)
-
     try:
+        user_input = req.text.strip()
+        if not user_input:
+            return {"reply": "", "tokens_generated": 0}
+
+        # Construir prompt
+        full_prompt = build_prompt(user_input)
+        tokens = [tokenizer_wrapper.bos_id()] + tokenizer_wrapper.encode(full_prompt)
+        input_ids = torch.tensor([tokens], device=DEVICE)
+
+        # Parámetros dinámicos según carga
+        dyn_max_tokens = req.max_tokens
+        dyn_temperature = req.temperature
+        
+        if ACTIVE_REQUESTS > 2:
+            dyn_max_tokens = min(dyn_max_tokens, 80)
+            dyn_temperature = max(0.5, dyn_temperature * 0.9)
+
+        # Generar
         with torch.no_grad():
-            output_ids = model.generate(
+            output_ids, eos_detected = model.generate(
                 input_ids,
                 max_new_tokens=dyn_max_tokens,
                 temperature=dyn_temperature,
                 top_k=req.top_k,
                 top_p=req.top_p,
-                repetition_penalty=req.repetition_penalty
+                repetition_penalty=req.repetition_penalty,
+                eos_token_id=tokenizer_wrapper.eos_id()
             )
 
+        # Extraer solo los tokens generados (excluyendo el prompt)
         gen_tokens = output_ids[0, len(tokens):].tolist()
         
+        # Filtrar tokens inválidos
         safe_tokens = [
             t for t in gen_tokens 
             if 0 <= t < VOCAB_SIZE and t != tokenizer_wrapper.eos_id()
         ]
         
-        response = tokenizer_wrapper.decode(safe_tokens).strip()
+        # Decodificar
+        raw_response = tokenizer_wrapper.decode(safe_tokens).strip()
         
-        if "###" in response:
-            response = response.split("###")[0].strip()
+        # Limpiar respuesta
+        clean_reply = clean_response(raw_response, user_input)
+        
+        # Si EOS no fue detectado y la respuesta parece incompleta, truncar
+        if not eos_detected and len(clean_reply) > 200:
+            # Buscar un punto final para truncar
+            last_period = clean_reply.rfind('.')
+            if last_period > 100:
+                clean_reply = clean_reply[:last_period + 1]
+        
+        # Eliminar frases sin sentido comunes
+        nonsense_phrases = [
+            "foompañances", "ciudadores", "mejtedon", "calportedon",
+            "rápidodcor", "rápidodarse", "miel", "baon", "domol"
+        ]
+        for phrase in nonsense_phrases:
+            clean_reply = clean_reply.replace(phrase, "")
+        
+        # Limpiar espacios dobles nuevamente
+        clean_reply = re.sub(r'\s+', ' ', clean_reply).strip()
+        
+        # Si la respuesta sigue siendo muy larga y no tiene puntos, cortar
+        if len(clean_reply) > 300 and '.' not in clean_reply[-50:]:
+            clean_reply = clean_reply[:250] + "..."
 
         return {
-            "reply": response,
+            "reply": clean_reply,
             "tokens_generated": len(safe_tokens),
-            "model": "MTP-1.1"
+            "model": "MTP-1.1",
+            "eos_detected": eos_detected
         }
     
     except Exception as e:
@@ -393,7 +495,7 @@ def model_info():
     }
 
 # ======================
-# INTERFAZ WEB (MODERNA DE MTP-3)
+# INTERFAZ WEB (MODERNA)
 # ======================
 @app.get("/", response_class=HTMLResponse)
 def chat_ui():
@@ -403,7 +505,7 @@ def chat_ui():
 <head>
 <meta charset="UTF-8">
 <meta name="viewport" content="width=device-width, initial-scale=1.0, maximum-scale=1.0, user-scalable=no">
-<title>MTP 1.1</title>
+<title>MTP 1.1 - Chat IA</title>
 <link rel="preconnect" href="https://fonts.googleapis.com">
 <link rel="preconnect" href="https://fonts.gstatic.com" crossorigin>
 <link href="https://fonts.googleapis.com/css2?family=Inter:wght@300;400;500;600&display=swap" rel="stylesheet">
@@ -415,8 +517,6 @@ def chat_ui():
     --text-primary: #e3e3e3;
     --text-secondary: #9aa0a6;
     --user-bubble: #282a2c;
-    --bot-actions-color: #c4c7c5;
-    --logo-url: url('https://i.postimg.cc/yxS54PF3/IMG-3082.jpg');
 }
 * { box-sizing: border-box; outline: none; -webkit-tap-highlight-color: transparent; }
 body {
@@ -452,10 +552,13 @@ header {
     width: 32px;
     height: 32px;
     border-radius: 50%;
-    background-image: var(--logo-url);
-    background-size: cover;
-    background-position: center;
-    border: 1px solid rgba(255,255,255,0.1);
+    background: linear-gradient(135deg, #4a9eff, #8a6eff);
+    display: flex;
+    align-items: center;
+    justify-content: center;
+    font-weight: bold;
+    font-size: 14px;
+    color: white;
 }
 .brand-text {
     font-weight: 500;
@@ -472,6 +575,28 @@ header {
     border-radius: 12px;
     font-weight: 600;
 }
+.status-badge {
+    font-size: 0.7rem;
+    background: rgba(76, 175, 80, 0.15);
+    color: #4caf50;
+    padding: 2px 8px;
+    border-radius: 12px;
+    font-weight: 500;
+    display: flex;
+    align-items: center;
+    gap: 6px;
+}
+.status-badge .dot {
+    width: 8px;
+    height: 8px;
+    background: #4caf50;
+    border-radius: 50%;
+    animation: pulse 1.5s infinite;
+}
+@keyframes pulse {
+    0%, 100% { opacity: 1; transform: scale(1); }
+    50% { opacity: 0.5; transform: scale(0.8); }
+}
 .chat-scroll {
     flex: 1;
     overflow-y: auto;
@@ -522,8 +647,13 @@ header {
     height: 34px;
     min-width: 34px;
     border-radius: 50%;
-    background-image: var(--logo-url);
-    background-size: cover;
+    background: linear-gradient(135deg, #4a9eff, #8a6eff);
+    display: flex;
+    align-items: center;
+    justify-content: center;
+    font-weight: bold;
+    font-size: 14px;
+    color: white;
     box-shadow: 0 2px 6px rgba(0,0,0,0.2);
 }
 .bot-actions {
@@ -590,6 +720,9 @@ header {
     font-family: inherit;
     padding: 10px 0;
 }
+#userInput::placeholder {
+    color: var(--text-secondary);
+}
 #mainBtn {
     background: white;
     color: black;
@@ -616,12 +749,6 @@ header {
     to { opacity: 1; transform: translateY(0); }
 }
 @keyframes blink { 0%, 100% { opacity: 1; } 50% { opacity: 0; } }
-@keyframes pulseAvatar {
-    0% { box-shadow: 0 0 0 0 rgba(74, 158, 255, 0.4); }
-    70% { box-shadow: 0 0 0 8px rgba(74, 158, 255, 0); }
-    100% { box-shadow: 0 0 0 0 rgba(74, 158, 255, 0); }
-}
-.pulsing { animation: pulseAvatar 1.5s infinite; }
 ::-webkit-scrollbar { width: 8px; }
 ::-webkit-scrollbar-track { background: transparent; }
 ::-webkit-scrollbar-thumb { background: #333; border-radius: 4px; }
@@ -630,18 +757,22 @@ header {
 <body>
 <header>
     <div class="brand-wrapper" onclick="location.reload()">
-        <div class="brand-logo"></div>
+        <div class="brand-logo">M</div>
         <div class="brand-text">
-            MTP <span class="version-badge">1.1</span>
+            MTP <span class="version-badge">2.5</span>
         </div>
     </div>
+    <div class="status-badge">
+        <span class="dot"></span>
+        <span id="statusText">Conectado</span>
+    </div>
 </header>
 <div id="chatScroll" class="chat-scroll">
     <div class="msg-row bot" style="animation-delay: 0.1s;">
-        <div class="bot-avatar"></div>
+        <div class="bot-avatar">M</div>
         <div class="msg-content-wrapper">
             <div class="msg-text">
-                ¡Hola! Soy MTP 1.1. ¿En qué puedo ayudarte hoy?
+                ¡Hola! Soy MTP 2.5 ¿En qué puedo ayudarte hoy?
             </div>
         </div>
     </div>
@@ -659,26 +790,34 @@ header {
 const chatScroll = document.getElementById('chatScroll');
 const userInput = document.getElementById('userInput');
 const mainBtn = document.getElementById('mainBtn');
+const statusText = document.getElementById('statusText');
 let isGenerating = false;
 let abortController = null;
 let typingTimeout = null;
 let lastUserPrompt = "";
+
 const ICON_SEND = `<svg width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2"><path d="M22 2L11 13M22 2l-7 20-4-9-9-4 20-7z"></path></svg>`;
 const ICON_STOP = `<svg width="14" height="14" viewBox="0 0 24 24" fill="currentColor" stroke="currentColor" stroke-width="0"><rect x="2" y="2" width="20" height="20" rx="4" ry="4"></rect></svg>`;
+
 mainBtn.innerHTML = ICON_SEND;
+
 function scrollToBottom() {
     chatScroll.scrollTop = chatScroll.scrollHeight;
 }
+
 function setBtnState(state) {
     if (state === 'sending') {
         mainBtn.innerHTML = ICON_STOP;
         isGenerating = true;
+        statusText.textContent = "Pensando...";
     } else {
         mainBtn.innerHTML = ICON_SEND;
         isGenerating = false;
         abortController = null;
+        statusText.textContent = "Conectado";
     }
 }
+
 function handleBtnClick() {
     if (isGenerating) {
         stopGeneration();
@@ -686,39 +825,44 @@ function handleBtnClick() {
         sendMessage();
     }
 }
+
 function stopGeneration() {
     if (abortController) abortController.abort();
     if (typingTimeout) clearTimeout(typingTimeout);
     const activeCursor = document.querySelector('.typing-cursor');
     if (activeCursor) activeCursor.classList.remove('typing-cursor');
-    const activeAvatar = document.querySelector('.pulsing');
-    if (activeAvatar) activeAvatar.classList.remove('pulsing');
     setBtnState('idle');
     userInput.focus();
 }
+
 async function sendMessage(textOverride = null) {
     const text = textOverride || userInput.value.trim();
     if (!text) return;
     lastUserPrompt = text;
+    
     if (!textOverride) {
         userInput.value = '';
         addMessage(text, 'user');
     }
+    
     setBtnState('sending');
     abortController = new AbortController();
+    
     const botRow = document.createElement('div');
     botRow.className = 'msg-row bot';
     const avatar = document.createElement('div');
-    avatar.className = 'bot-avatar pulsing'; 
+    avatar.className = 'bot-avatar';
+    avatar.textContent = 'M';
     const wrapper = document.createElement('div');
     wrapper.className = 'msg-content-wrapper';
     const msgText = document.createElement('div');
-    msgText.className = 'msg-text'; 
+    msgText.className = 'msg-text';
     wrapper.appendChild(msgText);
     botRow.appendChild(avatar);
     botRow.appendChild(wrapper);
     chatScroll.appendChild(botRow);
     scrollToBottom();
+    
     try {
         const response = await fetch('/generate', {
             method: 'POST',
@@ -726,11 +870,15 @@ async function sendMessage(textOverride = null) {
             body: JSON.stringify({ text: text }),
             signal: abortController.signal
         });
+        
         const data = await response.json();
-        if (!isGenerating) return; 
-        avatar.classList.remove('pulsing');
-        const reply = data.reply || "No entendí eso.";
+        
+        if (!isGenerating) return;
+        
+        const reply = data.reply || "Lo siento, no pude procesar tu solicitud.";
+        
         await typeWriter(msgText, reply);
+        
         if (isGenerating) {
             addActions(wrapper, reply);
             setBtnState('idle');
@@ -739,13 +887,13 @@ async function sendMessage(textOverride = null) {
         if (error.name === 'AbortError') {
             msgText.textContent += " [Detenido]";
         } else {
-            avatar.classList.remove('pulsing');
-            msgText.textContent = "Error de conexión.";
+            msgText.textContent = "Error de conexión. Intenta de nuevo.";
             msgText.style.color = "#ff8b8b";
-            setBtnState('idle');
         }
+        setBtnState('idle');
     }
 }
+
 function addMessage(text, sender) {
     const row = document.createElement('div');
     row.className = `msg-row ${sender}`;
@@ -756,10 +904,13 @@ function addMessage(text, sender) {
     chatScroll.appendChild(row);
     scrollToBottom();
 }
+
 function typeWriter(element, text, speed = 12) {
     return new Promise(resolve => {
         let i = 0;
+        element.textContent = '';
         element.classList.add('typing-cursor');
+        
         function type() {
             if (!isGenerating) {
                 element.classList.remove('typing-cursor');
@@ -779,30 +930,43 @@ function typeWriter(element, text, speed = 12) {
         type();
     });
 }
+
 function addActions(wrapperElement, textToCopy) {
     const actionsDiv = document.createElement('div');
     actionsDiv.className = 'bot-actions';
+    
     const copyBtn = document.createElement('button');
     copyBtn.className = 'action-btn';
     copyBtn.innerHTML = `<svg viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><rect x="9" y="9" width="13" height="13" rx="2" ry="2"></rect><path d="M5 15H4a2 2 0 0 1-2-2V4a2 2 0 0 1 2-2h9a2 2 0 0 1 2 2v1"></path></svg>`;
     copyBtn.onclick = () => {
         navigator.clipboard.writeText(textToCopy);
+        copyBtn.innerHTML = `<svg viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2"><path d="M20 6L9 17l-5-5"></path></svg>`;
+        setTimeout(() => {
+            copyBtn.innerHTML = `<svg viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><rect x="9" y="9" width="13" height="13" rx="2" ry="2"></rect><path d="M5 15H4a2 2 0 0 1-2-2V4a2 2 0 0 1 2-2h9a2 2 0 0 1 2 2v1"></path></svg>`;
+        }, 1500);
     };
+    
     const regenBtn = document.createElement('button');
     regenBtn.className = 'action-btn';
     regenBtn.innerHTML = `<svg viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><path d="M23 4v6h-6"></path><path d="M1 20v-6h6"></path><path d="M3.51 9a9 9 0 0 1 14.85-3.36L23 10M1 14l4.64 4.36A9 9 0 0 0 20.49 15"></path></svg>`;
     regenBtn.onclick = () => {
         sendMessage(lastUserPrompt);
     };
+    
     actionsDiv.appendChild(copyBtn);
     actionsDiv.appendChild(regenBtn);
     wrapperElement.appendChild(actionsDiv);
     requestAnimationFrame(() => actionsDiv.style.opacity = "1");
     scrollToBottom();
 }
+
 userInput.addEventListener('keydown', (e) => {
-    if (e.key === 'Enter') handleBtnClick();
+    if (e.key === 'Enter' && !e.shiftKey) {
+        e.preventDefault();
+        handleBtnClick();
+    }
 });
+
 window.onload = () => userInput.focus();
 </script>
 </body>