### Problem
The `forward` method in `modeling.py` line 646 incorrectly assigns `outputs.hidden_states` to the `hidden_states` field in its return object.
The correct tensor, `outputs.last_hidden_state`, is already computed and used internally to calculate the logits, but it is effectively discarded in the final return value.
### Solution
This commit corrects a single variable in the `return` statement. It changes `outputs.hidden_states` to `outputs.last_hidden_state`, ensuring the final hidden state tensor is correctly passed through.
This makes the model's output useful for feature extraction and analysis, fixing the intended behavior without any other functional changes.
