diff --git a/training/weight_mapping.py b/training/weight_mapping.py
index 1abad45..ae1a205 100644
--- a/training/weight_mapping.py
+++ b/training/weight_mapping.py
@@ -48,10 +48,9 @@ def vllm_to_hf_views(vllm_params: dict[str, torch.Tensor]
     hf_params = {}
 
     for name, tensor in vllm_params.items():
-        # vLLM and HF both use 'language_model.model.layers...' for Qwen3.5.
-        # HF checkpoint has 'model.' prefix but named_parameters() doesn't.
-        # Keep vLLM's names as-is — we'll match when loading into the HF model.
-        hf_name = name
+        # vLLM uses 'language_model.model.layers...' but HF's text model
+        # uses 'model.layers...'. Strip the 'language_model.' prefix.
+        hf_name = name.removeprefix('language_model.')
 
         # Split merged projections into HF-style separate weights
         if 'in_proj_qkvz' in name: