Fix `max_steps` documentation regarding the end-of-training condition (#27624)

* fix max_steps doc

* Update src/transformers/training_args.py [ci skip]
Co-authored-by: Steven Liu <59462357+stevhliu@users.noreply.github.com>

* propagate suggested change

---------
Co-authored-by: Steven Liu <59462357+stevhliu@users.noreply.github.com>

src/transformers/training_args.py

@@ -234,8 +234,8 @@ class TrainingArguments:
             the last epoch before stopping training).
         max_steps (`int`, *optional*, defaults to -1):
             If set to a positive number, the total number of training steps to perform. Overrides `num_train_epochs`.
-            In case of using a finite iterable dataset the training may stop before reaching the set number of steps
-            when all data is exhausted
+            For a finite dataset, training is reiterated through the dataset (if all data is exhausted) until
+            `max_steps` is reached.
         lr_scheduler_type (`str` or [`SchedulerType`], *optional*, defaults to `"linear"`):
             The scheduler type to use. See the documentation of [`SchedulerType`] for all possible values.
         lr_scheduler_kwargs ('dict', *optional*, defaults to {}):
@@ -2181,9 +2181,9 @@ class TrainingArguments:
                 Total number of training epochs to perform (if not an integer, will perform the decimal part percents
                 of the last epoch before stopping training).
             max_steps (`int`, *optional*, defaults to -1):
-                If set to a positive number, the total number of training steps to perform. Overrides
-                `num_train_epochs`. In case of using a finite iterable dataset the training may stop before reaching
-                the set number of steps when all data is exhausted.
+                If set to a positive number, the total number of training steps to perform. Overrides `num_train_epochs`.
+                For a finite dataset, training is reiterated through the dataset (if all data is exhausted) until
+                `max_steps` is reached.
             gradient_accumulation_steps (`int`, *optional*, defaults to 1):
                 Number of updates steps to accumulate the gradients for, before performing a backward/update pass.
 
@@ -2588,9 +2588,9 @@ class TrainingArguments:
                 Total number of training epochs to perform (if not an integer, will perform the decimal part percents
                 of the last epoch before stopping training).
             max_steps (`int`, *optional*, defaults to -1):
-                If set to a positive number, the total number of training steps to perform. Overrides
-                `num_train_epochs`. In case of using a finite iterable dataset the training may stop before reaching
-                the set number of steps when all data is exhausted.
+                If set to a positive number, the total number of training steps to perform. Overrides `num_train_epochs`.
+                For a finite dataset, training is reiterated through the dataset (if all data is exhausted) until
+                `max_steps` is reached.
             warmup_ratio (`float`, *optional*, defaults to 0.0):
                 Ratio of total training steps used for a linear warmup from 0 to `learning_rate`.
             warmup_steps (`int`, *optional*, defaults to 0):

src/transformers/training_args_tf.py

@@ -92,6 +92,8 @@ class TFTrainingArguments(TrainingArguments):
             Total number of training epochs to perform.
         max_steps (`int`, *optional*, defaults to -1):
             If set to a positive number, the total number of training steps to perform. Overrides `num_train_epochs`.
+            For a finite dataset, training is reiterated through the dataset (if all data is exhausted) until
+            `max_steps` is reached.
         warmup_ratio (`float`, *optional*, defaults to 0.0):
             Ratio of total training steps used for a linear warmup from 0 to `learning_rate`.
         warmup_steps (`int`, *optional*, defaults to 0):