diff --git a/finetuning/sft_12hz.py b/finetuning/sft_12hz.py
index 6ec80c8..c1f3f46 100644
--- a/finetuning/sft_12hz.py
+++ b/finetuning/sft_12hz.py
@@ -105,7 +105,7 @@ def train():
                 )
 
                 hidden_states = outputs.hidden_states[0][-1]
-                talker_hidden_states = hidden_states[codec_mask[:, 1:]]
+                talker_hidden_states = hidden_states[codec_mask[:, :-1]]
                 talker_codec_ids = codec_ids[codec_mask]
 
                 sub_talker_logits, sub_talker_loss = model.talker.forward_sub_talker_finetune(talker_codec_ids, talker_hidden_states)