| """Hugging Face configuration for the educational FineWeb decoder.""" | |
| from transformers import PreTrainedConfig | |
| class FineWebConfig(PreTrainedConfig): | |
| model_type = "fineweb_decoder" | |
| def __init__( | |
| self, | |
| vocab_size=16384, | |
| n_layers=8, | |
| d_model=256, | |
| n_heads=4, | |
| mlp_hidden=960, | |
| context_length=1024, | |
| rms_norm_eps=1e-6, | |
| **kwargs, | |
| ): | |
| kwargs.setdefault("tie_word_embeddings", True) | |
| kwargs.setdefault("use_cache", False) | |
| super().__init__(**kwargs) | |
| self.vocab_size = vocab_size | |
| self.n_layers = n_layers | |
| self.d_model = d_model | |
| self.n_heads = n_heads | |
| self.mlp_hidden = mlp_hidden | |
| self.context_length = context_length | |
| self.max_position_embeddings = context_length | |
| self.rms_norm_eps = rms_norm_eps | |
| self.hidden_size = d_model | |
| self.num_hidden_layers = n_layers | |
| self.num_attention_heads = n_heads | |