End of training

README.md

@@ -0,0 +1,54 @@
+---
+library_name: transformers
+tags:
+- generated_from_trainer
+model-index:
+- name: ecommerce-domain-24m
+  results: []
+---
+
+<!-- This model card has been generated automatically according to the information the Trainer had access to. You
+should probably proofread and complete it, then remove this comment. -->
+
+# ecommerce-domain-24m
+
+This model is a fine-tuned version of [](https://huggingface.co/) on the None dataset.
+
+## Model description
+
+More information needed
+
+## Intended uses & limitations
+
+More information needed
+
+## Training and evaluation data
+
+More information needed
+
+## Training procedure
+
+### Training hyperparameters
+
+The following hyperparameters were used during training:
+- learning_rate: 0.0003
+- train_batch_size: 32
+- eval_batch_size: 32
+- seed: 42
+- gradient_accumulation_steps: 4
+- total_train_batch_size: 128
+- optimizer: Use adamw_torch_fused with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
+- lr_scheduler_type: cosine
+- lr_scheduler_warmup_steps: 200
+- num_epochs: 3
+
+### Training results
+
+
+
+### Framework versions
+
+- Transformers 5.5.0
+- Pytorch 2.10.0+cu128
+- Datasets 4.3.0
+- Tokenizers 0.22.2

configuration.py

@@ -0,0 +1,101 @@
+"""
+DomainTransformer Configuration.
+
+HF-compatible config following Nubank nuFormer architecture choices:
+  - GPT-style causal decoder
+  - NoPE (no positional encoding) by default — Kazemnejad et al. 2023
+  - Pre-norm (LayerNorm before attention and FFN)
+  - Weight-tied embedding ↔ LM head
+  - Two reference sizes: 24M (d=512, L=6) and 330M (d=1024, L=24)
+"""
+
+from transformers import PretrainedConfig
+
+
+class DomainTransformerConfig(PretrainedConfig):
+    """Configuration for DomainTransformer causal language model.
+
+    This config produces a GPT-style decoder-only Transformer with:
+      - No positional encoding (NoPE) by default
+      - Pre-norm architecture (LayerNorm before attn/FFN)
+      - GELU activation in FFN
+      - Weight tying between token embeddings and LM head
+
+    Predefined sizes following Nubank nuFormer (arXiv:2507.23267):
+      - "24m":  6 layers, d_model=512,  8 heads, FFN=2048   (~24M params)
+      - "85m":  12 layers, d_model=768,  12 heads, FFN=3072  (~85M params)
+      - "330m": 24 layers, d_model=1024, 16 heads, FFN=4096  (~330M params)
+
+    Args:
+        vocab_size: Size of the token vocabulary.
+        hidden_size: Dimension of hidden representations (d_model).
+        num_hidden_layers: Number of transformer blocks.
+        num_attention_heads: Number of attention heads.
+        intermediate_size: FFN intermediate dimension (default: 4 * hidden_size).
+        hidden_act: Activation function in FFN.
+        hidden_dropout_prob: Dropout rate for embeddings and residual connections.
+        attention_probs_dropout_prob: Dropout rate for attention weights.
+        max_position_embeddings: Maximum sequence length (for buffer sizing, not PE).
+        initializer_range: Std for weight initialization (normal distribution).
+        layer_norm_eps: Epsilon for LayerNorm.
+        use_cache: Whether to return past key values for generation.
+        tie_word_embeddings: Whether to tie input/output embeddings.
+    """
+
+    model_type = "domain_transformer"
+
+    def __init__(
+        self,
+        vocab_size: int = 32000,
+        hidden_size: int = 512,
+        num_hidden_layers: int = 6,
+        num_attention_heads: int = 8,
+        intermediate_size: int = None,
+        hidden_act: str = "gelu",
+        hidden_dropout_prob: float = 0.0,
+        attention_probs_dropout_prob: float = 0.0,
+        max_position_embeddings: int = 2048,
+        initializer_range: float = 0.02,
+        layer_norm_eps: float = 1e-5,
+        use_cache: bool = True,
+        tie_word_embeddings: bool = True,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.intermediate_size = intermediate_size if intermediate_size is not None else 4 * hidden_size
+        self.hidden_act = hidden_act
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.max_position_embeddings = max_position_embeddings
+        self.initializer_range = initializer_range
+        self.layer_norm_eps = layer_norm_eps
+        self.use_cache = use_cache
+
+        assert hidden_size % num_attention_heads == 0, (
+            f"hidden_size ({hidden_size}) must be divisible by "
+            f"num_attention_heads ({num_attention_heads})"
+        )
+
+        super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs)
+
+    @classmethod
+    def from_preset(cls, name: str, vocab_size: int = 32000, **overrides) -> "DomainTransformerConfig":
+        """Create config from a named preset.
+
+        Presets:
+            "24m":  ~24M params  (6 layers, d=512, 8 heads)
+            "85m":  ~85M params  (12 layers, d=768, 12 heads)
+            "330m": ~330M params (24 layers, d=1024, 16 heads)
+        """
+        presets = {
+            "24m": dict(hidden_size=512, num_hidden_layers=6, num_attention_heads=8, intermediate_size=2048),
+            "85m": dict(hidden_size=768, num_hidden_layers=12, num_attention_heads=12, intermediate_size=3072),
+            "330m": dict(hidden_size=1024, num_hidden_layers=24, num_attention_heads=16, intermediate_size=4096),
+        }
+        if name not in presets:
+            raise ValueError(f"Unknown preset '{name}'. Available: {list(presets.keys())}")
+        params = {**presets[name], "vocab_size": vocab_size, **overrides}
+        return cls(**params)

modeling.py

@@ -0,0 +1,209 @@
+"""
+DomainTransformer Model — GPT-style causal decoder for domain token sequences.
+
+Architecture follows:
+  - NoPE (no positional encoding) — Kazemnejad et al. 2023 (arXiv:2305.19466)
+  - Pre-norm (LayerNorm before attention and FFN) — GPT-2 style
+  - F.scaled_dot_product_attention with is_causal=True — auto FlashAttention
+  - Weight tying between token embedding and LM head
+  - Scaled residual initialization: 1/sqrt(2*N_layers)
+
+Reference sizes (Nubank nuFormer, arXiv:2507.23267):
+  - 24M:  6 layers, d=512, 8 heads
+  - 330M: 24 layers, d=1024, 16 heads
+"""
+
+import math
+from typing import Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from transformers import PreTrainedModel
+from transformers.modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast
+
+from .configuration import DomainTransformerConfig
+
+
+class DomainTransformerAttention(nn.Module):
+    """Multi-head self-attention with NoPE.
+
+    Uses F.scaled_dot_product_attention for automatic FlashAttention/SDPA dispatch.
+    No positional encoding — causal masking via is_causal=True.
+    """
+
+    def __init__(self, config: DomainTransformerConfig):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.hidden_size // self.num_heads
+        self.scaling = self.head_dim ** -0.5
+
+        self.q_proj = nn.Linear(self.hidden_size, self.hidden_size, bias=False)
+        self.k_proj = nn.Linear(self.hidden_size, self.hidden_size, bias=False)
+        self.v_proj = nn.Linear(self.hidden_size, self.hidden_size, bias=False)
+        self.out_proj = nn.Linear(self.hidden_size, self.hidden_size, bias=False)
+        self.attn_dropout = config.attention_probs_dropout_prob
+
+    def forward(self, hidden_states: torch.Tensor, attention_mask: Optional[torch.Tensor] = None) -> torch.Tensor:
+        B, T, C = hidden_states.shape
+        q = self.q_proj(hidden_states).view(B, T, self.num_heads, self.head_dim).transpose(1, 2)
+        k = self.k_proj(hidden_states).view(B, T, self.num_heads, self.head_dim).transpose(1, 2)
+        v = self.v_proj(hidden_states).view(B, T, self.num_heads, self.head_dim).transpose(1, 2)
+
+        # Convert HF-style attention_mask (1=attend, 0=ignore, long) to SDPA format
+        sdpa_mask = None
+        use_causal = True
+        if attention_mask is not None:
+            sdpa_mask = attention_mask[:, None, None, :].to(dtype=q.dtype)
+            sdpa_mask = (1.0 - sdpa_mask) * torch.finfo(q.dtype).min
+            use_causal = False
+
+        attn_out = F.scaled_dot_product_attention(
+            q, k, v, attn_mask=sdpa_mask,
+            dropout_p=self.attn_dropout if self.training else 0.0,
+            is_causal=use_causal, scale=self.scaling,
+        )
+        attn_out = attn_out.transpose(1, 2).contiguous().reshape(B, T, C)
+        return self.out_proj(attn_out)
+
+
+class DomainTransformerMLP(nn.Module):
+    """Two-layer FFN with GELU activation (GPT-2 style)."""
+
+    def __init__(self, config: DomainTransformerConfig):
+        super().__init__()
+        self.up_proj = nn.Linear(config.hidden_size, config.intermediate_size, bias=True)
+        self.down_proj = nn.Linear(config.intermediate_size, config.hidden_size, bias=True)
+        self.act = nn.GELU(approximate="tanh")
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        return self.dropout(self.down_proj(self.act(self.up_proj(hidden_states))))
+
+
+class DomainTransformerBlock(nn.Module):
+    """Single transformer block with pre-norm architecture."""
+
+    def __init__(self, config: DomainTransformerConfig):
+        super().__init__()
+        self.ln_1 = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.attn = DomainTransformerAttention(config)
+        self.ln_2 = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.mlp = DomainTransformerMLP(config)
+
+    def forward(self, hidden_states: torch.Tensor, attention_mask: Optional[torch.Tensor] = None) -> torch.Tensor:
+        residual = hidden_states
+        hidden_states = self.attn(self.ln_1(hidden_states), attention_mask)
+        hidden_states = residual + hidden_states
+        residual = hidden_states
+        hidden_states = self.mlp(self.ln_2(hidden_states))
+        hidden_states = residual + hidden_states
+        return hidden_states
+
+
+class DomainTransformerPreTrainedModel(PreTrainedModel):
+    """Base class with weight initialization."""
+    config_class = DomainTransformerConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+
+    def _init_weights(self, module: nn.Module):
+        std = self.config.initializer_range
+        if isinstance(module, nn.Linear):
+            nn.init.normal_(module.weight, mean=0.0, std=std)
+            if module.bias is not None:
+                nn.init.zeros_(module.bias)
+        elif isinstance(module, nn.Embedding):
+            nn.init.normal_(module.weight, mean=0.0, std=std)
+            if module.padding_idx is not None:
+                nn.init.zeros_(module.weight[module.padding_idx])
+        elif isinstance(module, nn.LayerNorm):
+            nn.init.zeros_(module.bias)
+            nn.init.ones_(module.weight)
+
+
+class DomainTransformerModel(DomainTransformerPreTrainedModel):
+    """The bare DomainTransformer: embeddings + blocks + final layernorm."""
+
+    def __init__(self, config: DomainTransformerConfig):
+        super().__init__(config)
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size)
+        self.embed_dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.blocks = nn.ModuleList([DomainTransformerBlock(config) for _ in range(config.num_hidden_layers)])
+        self.ln_f = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.gradient_checkpointing = False
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+    def forward(self, input_ids=None, attention_mask=None, inputs_embeds=None, **kwargs):
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+        hidden_states = self.embed_dropout(inputs_embeds)
+        for block in self.blocks:
+            if self.gradient_checkpointing and self.training:
+                hidden_states = torch.utils.checkpoint.checkpoint(block, hidden_states, attention_mask, use_reentrant=False)
+            else:
+                hidden_states = block(hidden_states, attention_mask)
+        hidden_states = self.ln_f(hidden_states)
+        return BaseModelOutputWithPast(last_hidden_state=hidden_states)
+
+
+class DomainTransformerForCausalLM(DomainTransformerPreTrainedModel):
+    """DomainTransformer with a causal language modeling head.
+
+    The LM head is weight-tied with the token embedding layer.
+    Loss is computed via standard shifted cross-entropy.
+    """
+    _tied_weights_keys = {"lm_head.weight": "model.embed_tokens.weight"}
+
+    def __init__(self, config: DomainTransformerConfig):
+        super().__init__(config)
+        self.model = DomainTransformerModel(config)
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.embed_tokens = value
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def forward(self, input_ids=None, attention_mask=None, labels=None, inputs_embeds=None, **kwargs):
+        outputs = self.model(input_ids=input_ids, attention_mask=attention_mask, inputs_embeds=inputs_embeds)
+        hidden_states = outputs.last_hidden_state
+        logits = self.lm_head(hidden_states)
+
+        loss = None
+        if labels is not None:
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            loss = F.cross_entropy(shift_logits.view(-1, self.config.vocab_size), shift_labels.view(-1), ignore_index=-100)
+
+        return CausalLMOutputWithPast(loss=loss, logits=logits)
+
+    def get_user_embedding(self, input_ids, attention_mask=None):
+        """Extract user-level embedding from the last non-padding token."""
+        outputs = self.model(input_ids=input_ids, attention_mask=attention_mask)
+        hidden_states = outputs.last_hidden_state
+        if attention_mask is not None:
+            seq_lengths = attention_mask.sum(dim=1) - 1
+            batch_idx = torch.arange(hidden_states.size(0), device=hidden_states.device)
+            return hidden_states[batch_idx, seq_lengths]
+        else:
+            return hidden_states[:, -1, :]
+
+
+DomainTransformerConfig.register_for_auto_class()
+DomainTransformerForCausalLM.register_for_auto_class("AutoModelForCausalLM")