Before, when initiating a batch generation, the first sampled token was broadcasted to all rows. This change updates the logic to ensure that a unique token is sampled for each row, improving the diversity and independence of generated sequences within a batch.
This commit addresses the TODO item in scripts/base_eval.py by removing the dependency on the pandas library. Specifically, the pd.read_csv call used to load eval_meta_data.csv has been replaced with Python's built-in csv module.
