from transformers import PretrainedConfig class MultimodalYieldConfig(PretrainedConfig): """Config for the text + image (concat) multi-cutoff yield model. Normalization statistics are stored here so the model accepts raw Qwen3 text embeddings and raw DINOv3 image features at inference time. """ model_type = "yield-text-image" def __init__( self, text_model="Qwen/Qwen3-Embedding-0.6B", image_model="facebook/dinov3-vitb16-pretrain-lvd1689m", text_dim=1024, image_dim=768, K=22, text_mean=None, text_std=None, image_mean=None, image_std=None, y_mean=None, y_std=None, train_cutoffs=None, eval_cutoffs=None, d_model=64, nhead=4, num_layers=2, dim_ff=128, dropout=0.3, input_dropout=0.2, pool="last", **kwargs, ): super().__init__(**kwargs) self.text_model = text_model self.image_model = image_model self.text_dim = text_dim self.image_dim = image_dim self.K = K self.text_mean = text_mean self.text_std = text_std self.image_mean = image_mean self.image_std = image_std self.y_mean = y_mean self.y_std = y_std self.train_cutoffs = train_cutoffs self.eval_cutoffs = eval_cutoffs self.d_model = d_model self.nhead = nhead self.num_layers = num_layers self.dim_ff = dim_ff self.dropout = dropout self.input_dropout = input_dropout self.pool = pool