diff --git a/conversion/bert.py b/conversion/bert.py
index 8af6c53..6dfe463 100644
--- a/conversion/bert.py
+++ b/conversion/bert.py
@@ -28,6 +28,16 @@ class BertModel(TextModel):
                 cls_out_labels = None
         self.cls_out_labels = cls_out_labels
 
+        # Certains modèles déclarent architectures=["BertModel"] mais utilisent un
+        # tokenizer XLM-RoBERTa SentencePiece/Unigram (ex. intfloat/multilingual-e5-*,
+        # CamemBERT). Détecter ici pour router le vocab + chopper la matrice de positions
+        # exactement comme RobertaModel/NomicBertModel le font déjà.
+        self._tokenizer_is_xlmroberta = self._is_tokenizer_xlmroberta()
+        if self._tokenizer_is_xlmroberta:
+            self._xlmroberta_tokenizer_init()
+        else:
+            self._position_offset = None
+
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
         self.gguf_writer.add_causal_attention(False)
@@ -37,6 +47,8 @@ class BertModel(TextModel):
             self.gguf_writer.add_classifier_output_labels([v for k, v in sorted(self.cls_out_labels.items())])
 
     def set_vocab(self):
+        if getattr(self, "_tokenizer_is_xlmroberta", False):
+            return self._xlmroberta_set_vocab()
         tokens, toktypes, tokpre = self.get_vocab_base()
         self.vocab_size = len(tokens)
 
@@ -91,6 +103,11 @@ class BertModel(TextModel):
         return super().filter_tensors((name, gen))
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # XLM-R : chopper la matrice de positions de _position_offset (pad/bos),
+        # comme RobertaModel/NeoBert. No-op si _position_offset is None (BERT WordPiece).
+        if name == "embeddings.position_embeddings.weight" and getattr(self, "_position_offset", None) is not None:
+            data_torch = data_torch[self._position_offset:, :]
+
         if self.cls_out_labels:
             # For BertForSequenceClassification (direct projection layer)
             if name == "classifier.weight":
@@ -101,6 +118,20 @@ class BertModel(TextModel):
 
         yield from super().modify_tensors(data_torch, name, bid)
 
+    # Détection robuste du tokenizer XLM-RoBERTa (SentencePiece/Unigram).
+    # Présent aussi sur NomicBertModel ; remonté en base pour que tout modèle
+    # déclaré "BertModel" mais à tokenizer Unigram (e5 multilingue, CamemBERT) marche.
+    def _is_tokenizer_xlmroberta(self) -> bool:
+        tj = self.dir_model / "tokenizer.json"
+        if not tj.is_file():
+            return False
+        try:
+            with open(tj) as f:
+                toktyp = json.load(f)["model"]["type"]
+        except Exception:
+            return False
+        return toktyp == "Unigram"
+
     def _xlmroberta_tokenizer_init(self) -> None:
         # 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: