Upload folder using huggingface_hub

train/models/encoder.py

@@ -252,13 +252,17 @@ class VBertEncoder(BaseEncoder):
 
         # Replace text model embeddings with DecoupledEmbedding to match the
         # V-SPLADE weight layout (tok_embeddings.weight + additional_embedding.weight).
+        # The native ModernBertModel uses a plain nn.Embedding with the FULL vocab
+        # (50408). V-SPLADE splits this into main (50368) + additional (40).
         text_model = instance.encoder.model.text_model
         old_emb = text_model.get_input_embeddings()
+        additional_vocab = getattr(config, "additional_vocab_size", 40)
+        main_vocab = old_emb.num_embeddings - additional_vocab  # 50408 - 40 = 50368
         new_emb = DecoupledEmbedding(
-            num_embeddings=old_emb.num_embeddings,
-            num_additional_embeddings=getattr(config, "additional_vocab_size", 40),
+            num_embeddings=main_vocab,
+            num_additional_embeddings=additional_vocab,
             embedding_dim=old_emb.embedding_dim,
-            padding_idx=old_emb.padding_idx,
+            padding_idx=old_emb.padding_idx if old_emb.padding_idx is not None and old_emb.padding_idx < main_vocab else None,
         ).to(dtype=dtype)
         text_model.set_input_embeddings(new_emb)