diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py
index 7cd60c0a2bec3..482ba264dd3c5 100755
--- a/convert_hf_to_gguf.py
+++ b/convert_hf_to_gguf.py
@@ -2624,6 +2624,16 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter
 @Model.register("RobertaModel")
 class RobertaModel(BertModel):
     model_arch = gguf.MODEL_ARCH.BERT
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # we need the pad_token_id to know how to chop down position_embd matrix
+        if (pad_token_id := self.hparams.get("pad_token_id")) is not None:
+            self._position_offset = 1 + pad_token_id
+            if "max_position_embeddings" in self.hparams:
+                self.hparams["max_position_embeddings"] -= self._position_offset
+        else:
+            self._position_offset = None
 
     def set_vocab(self):
         """Support BPE tokenizers for roberta models"""
@@ -2641,6 +2651,19 @@ def set_vocab(self):
         else:
             return super().set_vocab()
 
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # if name starts with "roberta.", remove the prefix
+        # e.g. https://huggingface.co/BAAI/bge-reranker-v2-m3/tree/main
+        if name.startswith("roberta."):
+            name = name[8:]
+
+        # position embeddings start at pad_token_id + 1, so just chop down the weight tensor
+        if name == "embeddings.position_embeddings.weight":
+            if self._position_offset is not None:
+                data_torch = data_torch[self._position_offset:,:]
+
+        return super().modify_tensors(data_torch, name, bid)
+
 
 @Model.register("NomicBertModel")
 class NomicBertModel(BertModel):