fix: fix length loglikelihood

biotransformers/wrappers/transformers_wrappers.py

@@ -306,11 +306,11 @@ def compute_logits(
 
         # Remove padded logits
         logits = [
-            torch.from_numpy(logit.numpy().transpose()[:, :length].transpose())
+            torch.from_numpy(logit.numpy().transpose()[:, 1:length].transpose())
             for logit, length in zip(list(logits), lengths)
         ]
         labels = [
-            torch.from_numpy(label.numpy().transpose()[:, :length].transpose())
+            torch.from_numpy(label.numpy().transpose()[:, 1:length].transpose())
             for label, length in zip(list(labels), lengths)
         ]
 
@@ -377,7 +377,7 @@ def compute_probabilities(
         # Remove padded logits
         # Use transpose so that function works for MSA and sequence
         logits = [
-            torch.from_numpy(logit.numpy().transpose()[:, :length].transpose())
+            torch.from_numpy(logit.numpy().transpose()[:, 1:length].transpose())
             for logit, length in zip(list(logits), lengths)
         ]
         # Set to -inf logits that correspond to tokens that are not in tokens list
@@ -745,7 +745,7 @@ def finetune(
             random_token_prob,
             toks_per_batch,
             extra_toks_per_seq,
-            validation=False
+            validation=False,
         )
 
         if self._num_gpus == 0:
@@ -787,7 +787,7 @@ def finetune(
                 save_name = self._save_model(save_path, lightning_model)
         else:
             save_name = self._save_model(save_path, lightning_model)
-            
+
         # Load new model
         self._language_model._load_model(save_name)
         log.info("Training completed.")