Upload 4 files

config.yaml

@@ -0,0 +1,64 @@
+# MTP Mini - Configuración Mejorada para Generación Coherente
+
+model:
+  vocab_size: 4000
+  d_model: 512              # Aumentado para más capacidad
+  n_layers: 8               # Más capas
+  n_heads: 8                # Más cabezas de atención
+  d_ff: 2048                # 4x d_model
+  max_seq_len: 512          # Contexto más largo
+  dropout: 0.2              # Más dropout para evitar overfitting
+  use_swiglu: true          # Activación mejorada
+
+training:
+  batch_size: 4             # Batch más pequeño para corpus pequeño
+  accumulation_steps: 4     # Effective batch = 16
+  epochs: 20                # MENOS épocas
+  learning_rate: 0.0003     # LR más alto para convergencia rápida
+  min_lr: 0.00001
+  weight_decay: 0.1         # MÁS weight decay para regularización
+  max_grad_norm: 1.0        
+  num_threads: 4
+  save_every: 5
+  
+  # Early stopping
+  patience: 5               # Parar si no mejora en 5 epochs
+  min_delta: 0.001          # Mejora mínima requerida
+  
+  # Learning rate schedule
+  warmup_steps: 100
+  use_lr_scheduler: true
+  
+  # Regularización adicional
+  label_smoothing: 0.1
+  use_eos_loss_weight: true # Dar más peso al token EOS
+
+data:
+  corpus_path: corpus/mtp_mini_corpus.jsonl
+  min_text_length: 50       # Textos más largos
+  max_text_length: 2000     # Permitir respuestas largas
+  validation_split: 0.15
+  
+  # Augmentación de datos
+  use_augmentation: true
+  augmentation_prob: 0.3
+
+generation:
+  # Parámetros de generación mejorados
+  default_max_tokens: 150
+  default_temperature: 0.7
+  default_top_k: 40
+  default_top_p: 0.92
+  default_repetition_penalty: 1.15
+  min_response_length: 20
+  use_length_penalty: true
+  
+  # Control de coherencia
+  use_perplexity_filter: true
+  max_perplexity: 100.0
+  
+  # Stop sequences
+  stop_sequences:
+    - "###"
+    - "\n\n\n"
+    - "Instrucción:"

dataset.py

@@ -0,0 +1,98 @@
+import torch
+from torch.utils.data import Dataset
+import json
+import random
+
+
+class MTPDataset(Dataset):
+    """Dataset mejorado con augmentación de datos"""
+    
+    def __init__(self, corpus_path, tokenizer, max_seq_len=512, 
+                 use_augmentation=False, augmentation_prob=0.3):
+        self.tokenizer = tokenizer
+        self.max_seq_len = max_seq_len
+        self.use_augmentation = use_augmentation
+        self.augmentation_prob = augmentation_prob
+        self.data = []
+        
+        # Load corpus
+        with open(corpus_path, 'r', encoding='utf-8') as f:
+            for line in f:
+                entry = json.loads(line)
+                if 'instruction' in entry and 'response' in entry:
+                    self.data.append(entry)
+        
+        print(f"✓ Loaded {len(self.data)} examples from corpus")
+        if use_augmentation:
+            print(f"✓ Data augmentation enabled (prob={augmentation_prob})")
+    
+    def __len__(self):
+        return len(self.data)
+    
+    def augment_text(self, text):
+        """Augmentación simple de texto"""
+        if not self.use_augmentation or random.random() > self.augmentation_prob:
+            return text
+        
+        # Variación 1: Agregar espacios aleatorios (simula variaciones en formato)
+        if random.random() < 0.3:
+            text = text.strip()
+        
+        # Variación 2: Cambiar puntuación final
+        if random.random() < 0.2:
+            if text.endswith('.'):
+                text = text[:-1]
+            elif not text.endswith(('.', '!', '?')):
+                text = text + '.'
+        
+        return text
+    
+    def __getitem__(self, idx):
+        entry = self.data[idx]
+        
+        instruction = entry['instruction']
+        response = entry['response']
+        
+        # Aplicar augmentación
+        instruction = self.augment_text(instruction)
+        response = self.augment_text(response)
+        
+        # Formato mejorado
+        full_text = f"### Instrucción:\n{instruction}\n\n### Respuesta:\n{response}"
+        
+        # Tokenize
+        tokens = self.tokenizer.encode(full_text)
+        
+        # Add BOS and EOS
+        tokens = [self.tokenizer.bos_id()] + tokens + [self.tokenizer.eos_id()]
+        
+        # Truncate if too long
+        if len(tokens) > self.max_seq_len:
+            # Truncar manteniendo BOS y EOS
+            tokens = [tokens[0]] + tokens[1:self.max_seq_len-1] + [self.tokenizer.eos_id()]
+        
+        # Convert to tensor
+        input_ids = torch.tensor(tokens[:-1], dtype=torch.long)
+        target_ids = torch.tensor(tokens[1:], dtype=torch.long)
+        
+        return input_ids, target_ids
+
+
+def collate_fn(batch, pad_id=0):
+    """Custom collate function con padding inteligente"""
+    input_ids = [item[0] for item in batch]
+    target_ids = [item[1] for item in batch]
+    
+    # Find max length in batch
+    max_len = max(len(ids) for ids in input_ids)
+    
+    # Pad sequences
+    input_ids_padded = []
+    target_ids_padded = []
+    
+    for inp, tgt in zip(input_ids, target_ids):
+        pad_len = max_len - len(inp)
+        input_ids_padded.append(torch.cat([inp, torch.full((pad_len,), pad_id, dtype=torch.long)]))
+        target_ids_padded.append(torch.cat([tgt, torch.full((pad_len,), pad_id, dtype=torch.long)]))
+    
+    return torch.stack(input_ids_padded), torch.stack(target_ids_padded)

model.py

@@ -0,0 +1,291 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import math
+
+
+class RotaryPositionalEmbedding(nn.Module):
+    """RoPE - Rotary Position Embedding"""
+    
+    def __init__(self, dim, max_seq_len=2048, base=10000):
+        super().__init__()
+        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float() / dim))
+        self.register_buffer('inv_freq', inv_freq)
+        self.max_seq_len = max_seq_len
+        
+    def forward(self, seq_len, device):
+        t = torch.arange(seq_len, device=device).type_as(self.inv_freq)
+        freqs = torch.einsum('i,j->ij', t, self.inv_freq)
+        emb = torch.cat((freqs, freqs), dim=-1)
+        return emb.cos(), emb.sin()
+
+
+def apply_rotary_pos_emb(q, k, cos, sin):
+    """Aplica RoPE a queries y keys"""
+    def rotate_half(x):
+        x1, x2 = x.chunk(2, dim=-1)
+        return torch.cat((-x2, x1), dim=-1)
+    
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+class MultiHeadSelfAttention(nn.Module):
+    """Multi-Head Self-Attention con RoPE y optimizaciones"""
+    
+    def __init__(self, d_model, n_heads, dropout=0.1, max_seq_len=2048):
+        super().__init__()
+        assert d_model % n_heads == 0
+        
+        self.d_model = d_model
+        self.n_heads = n_heads
+        self.d_k = d_model // n_heads
+        
+        self.q_linear = nn.Linear(d_model, d_model, bias=False)
+        self.k_linear = nn.Linear(d_model, d_model, bias=False)
+        self.v_linear = nn.Linear(d_model, d_model, bias=False)
+        self.out_linear = nn.Linear(d_model, d_model, bias=False)
+        
+        self.dropout = nn.Dropout(dropout)
+        self.attn_dropout = nn.Dropout(dropout)
+        self.rope = RotaryPositionalEmbedding(self.d_k, max_seq_len)
+        
+        self.flash = hasattr(torch.nn.functional, 'scaled_dot_product_attention')
+        
+    def forward(self, x, mask=None):
+        batch_size, seq_len, d_model = x.size()
+        
+        Q = self.q_linear(x).view(batch_size, seq_len, self.n_heads, self.d_k).transpose(1, 2)
+        K = self.k_linear(x).view(batch_size, seq_len, self.n_heads, self.d_k).transpose(1, 2)
+        V = self.v_linear(x).view(batch_size, seq_len, self.n_heads, self.d_k).transpose(1, 2)
+        
+        cos, sin = self.rope(seq_len, x.device)
+        cos = cos[None, None, :, :]
+        sin = sin[None, None, :, :]
+        Q, K = apply_rotary_pos_emb(Q, K, cos, sin)
+        
+        if self.flash and mask is None:
+            context = F.scaled_dot_product_attention(
+                Q, K, V, 
+                attn_mask=None,
+                dropout_p=self.dropout.p if self.training else 0.0,
+                is_causal=True
+            )
+        else:
+            scores = torch.matmul(Q, K.transpose(-2, -1)) / math.sqrt(self.d_k)
+            if mask is not None:
+                scores = scores.masked_fill(mask == 0, float('-inf'))
+            attn_weights = F.softmax(scores, dim=-1)
+            attn_weights = self.attn_dropout(attn_weights)
+            context = torch.matmul(attn_weights, V)
+        
+        context = context.transpose(1, 2).contiguous().view(batch_size, seq_len, d_model)
+        output = self.out_linear(context)
+        return self.dropout(output)
+
+
+class SwiGLU(nn.Module):
+    """SwiGLU activation"""
+    
+    def __init__(self, d_model, d_ff, dropout=0.1):
+        super().__init__()
+        self.w1 = nn.Linear(d_model, d_ff, bias=False)
+        self.w2 = nn.Linear(d_ff, d_model, bias=False)
+        self.w3 = nn.Linear(d_model, d_ff, bias=False)
+        self.dropout = nn.Dropout(dropout)
+        
+    def forward(self, x):
+        return self.w2(self.dropout(F.silu(self.w1(x)) * self.w3(x)))
+
+
+class FeedForward(nn.Module):
+    """Feed-Forward estándar"""
+    
+    def __init__(self, d_model, d_ff, dropout=0.1):
+        super().__init__()
+        self.linear1 = nn.Linear(d_model, d_ff)
+        self.linear2 = nn.Linear(d_ff, d_model)
+        self.dropout = nn.Dropout(dropout)
+        
+    def forward(self, x):
+        return self.linear2(self.dropout(F.gelu(self.linear1(x))))
+
+
+class RMSNorm(nn.Module):
+    """RMSNorm"""
+    
+    def __init__(self, dim, eps=1e-6):
+        super().__init__()
+        self.eps = eps
+        self.weight = nn.Parameter(torch.ones(dim))
+        
+    def forward(self, x):
+        norm = torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
+        return x * norm * self.weight
+
+
+class TransformerBlock(nn.Module):
+    """Transformer Block mejorado"""
+    
+    def __init__(self, d_model, n_heads, d_ff, dropout=0.1, max_seq_len=2048, use_swiglu=True):
+        super().__init__()
+        self.attention = MultiHeadSelfAttention(d_model, n_heads, dropout, max_seq_len)
+        
+        if use_swiglu:
+            self.feed_forward = SwiGLU(d_model, d_ff, dropout)
+        else:
+            self.feed_forward = FeedForward(d_model, d_ff, dropout)
+        
+        self.norm1 = RMSNorm(d_model)
+        self.norm2 = RMSNorm(d_model)
+        self.dropout = nn.Dropout(dropout)
+        
+    def forward(self, x, mask=None):
+        x = x + self.attention(self.norm1(x), mask)
+        x = x + self.feed_forward(self.norm2(x))
+        return x
+
+
+class MTPMiniModel(nn.Module):
+    """MTP Mini - Modelo mejorado para generación coherente"""
+    
+    def __init__(self, vocab_size, d_model=512, n_layers=8, n_heads=8, 
+                 d_ff=2048, max_seq_len=512, dropout=0.2, use_swiglu=True):
+        super().__init__()
+        
+        self.vocab_size = vocab_size
+        self.d_model = d_model
+        self.max_seq_len = max_seq_len
+        
+        self.token_embedding = nn.Embedding(vocab_size, d_model)
+        self.dropout = nn.Dropout(dropout)
+        
+        self.blocks = nn.ModuleList([
+            TransformerBlock(d_model, n_heads, d_ff, dropout, max_seq_len, use_swiglu)
+            for _ in range(n_layers)
+        ])
+        
+        self.norm_f = RMSNorm(d_model)
+        self.lm_head = nn.Linear(d_model, vocab_size, bias=False)
+        
+        # Weight tying
+        self.lm_head.weight = self.token_embedding.weight
+        
+        self.apply(self._init_weights)
+        
+    def _init_weights(self, module):
+        if isinstance(module, nn.Linear):
+            torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)
+            if module.bias is not None:
+                torch.nn.init.zeros_(module.bias)
+        elif isinstance(module, nn.Embedding):
+            torch.nn.init.normal_(module.weight, mean=0.0, std=0.02)
+    
+    def forward(self, input_ids, targets=None, use_eos_weight=False):
+        batch_size, seq_len = input_ids.size()
+        
+        mask = torch.tril(torch.ones(seq_len, seq_len, device=input_ids.device)).view(1, 1, seq_len, seq_len)
+        
+        x = self.dropout(self.token_embedding(input_ids))
+        
+        for block in self.blocks:
+            x = block(x, mask)
+        
+        x = self.norm_f(x)
+        logits = self.lm_head(x)
+        
+        loss = None
+        if targets is not None:
+            if use_eos_weight:
+                # Dar más peso al token EOS para aprender a terminar
+                weights = torch.ones(self.vocab_size, device=logits.device)
+                weights[3] = 2.0  # EOS token
+                loss = F.cross_entropy(
+                    logits.view(-1, self.vocab_size), 
+                    targets.view(-1),
+                    weight=weights,
+                    label_smoothing=0.1
+                )
+            else:
+                loss = F.cross_entropy(
+                    logits.view(-1, self.vocab_size), 
+                    targets.view(-1),
+                    label_smoothing=0.1
+                )
+        
+        return logits, loss
+    
+    def generate(self, input_ids, max_new_tokens=150, temperature=0.7, 
+                 top_k=40, top_p=0.92, repetition_penalty=1.15,
+                 min_length=20, eos_token_id=3, stop_sequences=None):
+        """Generación mejorada con control de longitud y coherencia"""
+        self.eval()
+        
+        generated = input_ids.clone()
+        generated_text_tokens = 0
+        
+        with torch.no_grad():
+            for step in range(max_new_tokens):
+                input_ids_cond = generated if generated.size(1) <= self.max_seq_len else generated[:, -self.max_seq_len:]
+                
+                logits, _ = self(input_ids_cond)
+                logits = logits[:, -1, :].clone()
+                
+                # Repetition penalty mejorado
+                if repetition_penalty != 1.0:
+                    for token_id in set(generated[0].tolist()):
+                        if logits[0, token_id] < 0:
+                            logits[0, token_id] *= repetition_penalty
+                        else:
+                            logits[0, token_id] /= repetition_penalty
+                
+                # Penalizar tokens repetidos recientes más fuertemente
+                if generated.size(1) > 10:
+                    recent_tokens = generated[0, -10:].tolist()
+                    for token_id in set(recent_tokens):
+                        count = recent_tokens.count(token_id)
+                        if count > 2:
+                            logits[0, token_id] -= count * 2.0
+                
+                # No permitir EOS hasta longitud mínima
+                if generated_text_tokens < min_length:
+                    logits[0, eos_token_id] = float('-inf')
+                else:
+                    # Aumentar probabilidad de EOS gradualmente
+                    eos_boost = (generated_text_tokens - min_length) * 0.1
+                    logits[0, eos_token_id] += eos_boost
+                
+                # Temperature
+                logits = logits / temperature
+                
+                # Top-k
+                if top_k > 0:
+                    v, _ = torch.topk(logits, min(top_k, logits.size(-1)))
+                    logits[logits < v[:, [-1]]] = float('-inf')
+                
+                # Top-p (nucleus)
+                if top_p < 1.0:
+                    sorted_logits, sorted_indices = torch.sort(logits, descending=True)
+                    cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
+                    sorted_indices_to_remove = cumulative_probs > top_p
+                    sorted_indices_to_remove[:, 1:] = sorted_indices_to_remove[:, :-1].clone()
+                    sorted_indices_to_remove[:, 0] = 0
+                    indices_to_remove = sorted_indices_to_remove.scatter(1, sorted_indices, sorted_indices_to_remove)
+                    logits[indices_to_remove] = float('-inf')
+                
+                # Sample
+                probs = F.softmax(logits, dim=-1)
+                next_token = torch.multinomial(probs, num_samples=1)
+                
+                # Check for EOS
+                if next_token.item() == eos_token_id and generated_text_tokens >= min_length:
+                    break
+                
+                generated = torch.cat([generated, next_token], dim=1)
+                generated_text_tokens += 1
+        
+        return generated
+    
+    def count_parameters(self):
+        return sum(p.numel() for p in self.parameters() if p.requires_grad)

tokenizer.py

@@ -0,0 +1,138 @@
+import sentencepiece as spm
+import os
+import json
+
+
+class MTPTokenizer:
+    """Tokenizer using SentencePiece BPE"""
+    
+    def __init__(self, model_path=None):
+        self.sp = None
+        self.model_path = model_path
+        
+        if model_path and os.path.exists(model_path):
+            self.load(model_path)
+    
+    def train(self, corpus_path, vocab_size=4000, model_prefix='mtp_tokenizer'):
+        """Train SentencePiece BPE tokenizer on corpus"""
+        
+        # Extract text from JSONL corpus
+        texts = []
+        with open(corpus_path, 'r', encoding='utf-8') as f:
+            for line in f:
+                data = json.loads(line)
+                if 'instruction' in data:
+                    texts.append(data['instruction'])
+                if 'response' in data:
+                    texts.append(data['response'])
+        
+        # Save temporary text file
+        temp_file = 'temp_corpus.txt'
+        with open(temp_file, 'w', encoding='utf-8') as f:
+            f.write('\n'.join(texts))
+        
+        # Calculate optimal vocab size based on corpus
+        total_chars = sum(len(text) for text in texts)
+        max_vocab = min(vocab_size, int(total_chars * 0.15))  # Heuristic: ~15% of chars
+        
+        print(f"   → Corpus stats: {len(texts)} texts, {total_chars} characters")
+        print(f"   → Adjusted vocab size: {max_vocab} (requested: {vocab_size})")
+        
+        # Train SentencePiece with adjusted parameters
+        try:
+            spm.SentencePieceTrainer.train(
+                input=temp_file,
+                model_prefix=model_prefix,
+                vocab_size=max_vocab,
+                model_type='bpe',
+                pad_id=0,
+                unk_id=1,
+                bos_id=2,
+                eos_id=3,
+                character_coverage=1.0,
+                normalization_rule_name='identity',
+                num_threads=4,
+                split_digits=True,
+                allow_whitespace_only_pieces=False,
+                byte_fallback=False,
+                max_sentencepiece_length=16
+            )
+        except RuntimeError as e:
+            if "Vocabulary size too high" in str(e):
+                # Extract suggested max from error and retry
+                import re
+                match = re.search(r'value <= (\d+)', str(e))
+                if match:
+                    suggested_max = int(match.group(1))
+                    print(f"   → Retrying with vocab size: {suggested_max}")
+                    spm.SentencePieceTrainer.train(
+                        input=temp_file,
+                        model_prefix=model_prefix,
+                        vocab_size=suggested_max,
+                        model_type='bpe',
+                        pad_id=0,
+                        unk_id=1,
+                        bos_id=2,
+                        eos_id=3,
+                        character_coverage=1.0,
+                        normalization_rule_name='identity',
+                        num_threads=4,
+                        split_digits=True,
+                        allow_whitespace_only_pieces=False,
+                        byte_fallback=False,
+                        max_sentencepiece_length=16
+                    )
+                else:
+                    raise
+            else:
+                raise
+        
+        # Clean up
+        os.remove(temp_file)
+        
+        # Load the trained model
+        self.model_path = f"{model_prefix}.model"
+        self.load(self.model_path)
+        
+        print(f"✓ Tokenizer trained: {self.vocab_size()} tokens")
+        print(f"✓ Model saved: {self.model_path}")
+    
+    def load(self, model_path):
+        """Load trained tokenizer"""
+        self.sp = spm.SentencePieceProcessor()
+        self.sp.load(model_path)
+        self.model_path = model_path
+    
+    def encode(self, text):
+        """Encode text to token IDs"""
+        if self.sp is None:
+            raise ValueError("Tokenizer not loaded. Train or load a model first.")
+        return self.sp.encode_as_ids(text)
+    
+    def decode(self, ids):
+        """Decode token IDs to text"""
+        if self.sp is None:
+            raise ValueError("Tokenizer not loaded. Train or load a model first.")
+        return self.sp.decode_ids(ids)
+    
+    def vocab_size(self):
+        """Get vocabulary size"""
+        if self.sp is None:
+            return 0
+        return self.sp.get_piece_size()
+    
+    def bos_id(self):
+        """Beginning of sentence token ID"""
+        return self.sp.bos_id()
+    
+    def eos_id(self):
+        """End of sentence token ID"""
+        return self.sp.eos_id()
+    
+    def pad_id(self):
+        """Padding token ID"""
+        return self.sp.pad_id()
+    
+    def unk_id(self):
+        """Unknown token ID"""
+        return self.sp.unk_id()