🚨🚨🚨Deprecate `evaluation_strategy` to `eval_strategy`🚨🚨🚨 (#30190)

* Alias

* Note alias

* Tests and src

* Rest

* Clean

* Change typing?

* Fix tests

* Deprecation versions

examples/research_projects/wav2vec2/finetune_large_xlsr_53_arabic_speech_corpus.sh

@@ -5,7 +5,7 @@ python run_asr.py \
 --per_device_train_batch_size="1" \
 --per_device_eval_batch_size="1" \
 --gradient_accumulation_steps="8" \
---evaluation_strategy="steps" \
+--eval_strategy="steps" \
 --save_steps="500" \
 --eval_steps="100" \
 --logging_steps="50" \

examples/research_projects/wav2vec2/finetune_wav2vec2_xlsr_turkish.sh

@@ -6,7 +6,7 @@ python run_common_voice.py \
     --overwrite_output_dir \
     --num_train_epochs="5" \
     --per_device_train_batch_size="16" \
-    --evaluation_strategy="steps" \
+    --eval_strategy="steps" \
     --learning_rate="3e-4" \
     --warmup_steps="500" \
     --fp16 \

examples/research_projects/wav2vec2/test_wav2vec2_deepspeed.py

@@ -161,7 +161,7 @@ class TestDeepSpeedWav2Vec2(TestCasePlus):
             --num_train_epochs {str(num_train_epochs)}
             --per_device_train_batch_size 2
             --per_device_eval_batch_size 2
-            --evaluation_strategy steps
+            --eval_strategy steps
             --learning_rate 5e-4
             --warmup_steps 8
             --orthography timit

examples/research_projects/xtreme-s/README.md

@@ -90,7 +90,7 @@ python -m torch.distributed.launch \
     --gradient_accumulation_steps=2 \
     --learning_rate="3e-4" \
     --warmup_steps=3000 \
-    --evaluation_strategy="steps" \
+    --eval_strategy="steps" \
     --max_duration_in_seconds=20 \
     --save_steps=500 \
     --eval_steps=500 \
@@ -134,7 +134,7 @@ python -m torch.distributed.launch \
     --gradient_accumulation_steps=1 \
     --learning_rate="3e-4" \
     --warmup_steps=1500 \
-    --evaluation_strategy="steps" \
+    --eval_strategy="steps" \
     --max_duration_in_seconds=30 \
     --save_steps=200 \
     --eval_steps=200 \

examples/tensorflow/image-classification/README.md

@@ -45,7 +45,7 @@ python run_image_classification.py \
     --per_device_eval_batch_size 8 \
     --logging_strategy steps \
     --logging_steps 10 \
-    --evaluation_strategy epoch \
+    --eval_strategy epoch \
     --save_strategy epoch \
     --load_best_model_at_end True \
     --save_total_limit 3 \

src/transformers/integrations/integration_utils.py

@@ -320,13 +320,13 @@ def run_hp_search_ray(trainer, n_trials: int, direction: str, **kwargs) -> BestR
         # Check for `do_eval` and `eval_during_training` for schedulers that require intermediate reporting.
         if isinstance(
             kwargs["scheduler"], (ASHAScheduler, MedianStoppingRule, HyperBandForBOHB, PopulationBasedTraining)
-        ) and (not trainer.args.do_eval or trainer.args.evaluation_strategy == IntervalStrategy.NO):
+        ) and (not trainer.args.do_eval or trainer.args.eval_strategy == IntervalStrategy.NO):
             raise RuntimeError(
                 "You are using {cls} as a scheduler but you haven't enabled evaluation during training. "
                 "This means your trials will not report intermediate results to Ray Tune, and "
                 "can thus not be stopped early or used to exploit other trials parameters. "
                 "If this is what you want, do not use {cls}. If you would like to use {cls}, "
-                "make sure you pass `do_eval=True` and `evaluation_strategy='steps'` in the "
+                "make sure you pass `do_eval=True` and `eval_strategy='steps'` in the "
                 "Trainer `args`.".format(cls=type(kwargs["scheduler"]).__name__)
             )
 

src/transformers/trainer_callback.py

@@ -444,7 +444,7 @@ class DefaultFlowCallback(TrainerCallback):
 
         # Evaluate
         if (
-            args.evaluation_strategy == IntervalStrategy.STEPS
+            args.eval_strategy == IntervalStrategy.STEPS
             and state.global_step % state.eval_steps == 0
             and args.eval_delay <= state.global_step
         ):
@@ -470,7 +470,7 @@ class DefaultFlowCallback(TrainerCallback):
             control.should_log = True
 
         # Evaluate
-        if args.evaluation_strategy == IntervalStrategy.EPOCH and args.eval_delay <= state.epoch:
+        if args.eval_strategy == IntervalStrategy.EPOCH and args.eval_delay <= state.epoch:
             control.should_evaluate = True
 
         # Save
@@ -586,7 +586,7 @@ class EarlyStoppingCallback(TrainerCallback):
             args.metric_for_best_model is not None
         ), "EarlyStoppingCallback requires metric_for_best_model is defined"
         assert (
-            args.evaluation_strategy != IntervalStrategy.NO
+            args.eval_strategy != IntervalStrategy.NO
         ), "EarlyStoppingCallback requires IntervalStrategy of steps or epoch"
 
     def on_evaluate(self, args, state, control, metrics, **kwargs):

src/transformers/training_args.py

@@ -226,7 +226,7 @@ class TrainingArguments:
             by your training/evaluation scripts instead. See the [example
             scripts](https://github.com/huggingface/transformers/tree/main/examples) for more details.
         do_eval (`bool`, *optional*):
-            Whether to run evaluation on the validation set or not. Will be set to `True` if `evaluation_strategy` is
+            Whether to run evaluation on the validation set or not. Will be set to `True` if `eval_strategy` is
             different from `"no"`. This argument is not directly used by [`Trainer`], it's intended to be used by your
             training/evaluation scripts instead. See the [example
             scripts](https://github.com/huggingface/transformers/tree/main/examples) for more details.
@@ -234,7 +234,7 @@ class TrainingArguments:
             Whether to run predictions on the test set or not. This argument is not directly used by [`Trainer`], it's
             intended to be used by your training/evaluation scripts instead. See the [example
             scripts](https://github.com/huggingface/transformers/tree/main/examples) for more details.
-        evaluation_strategy (`str` or [`~trainer_utils.IntervalStrategy`], *optional*, defaults to `"no"`):
+        eval_strategy (`str` or [`~trainer_utils.IntervalStrategy`], *optional*, defaults to `"no"`):
             The evaluation strategy to adopt during training. Possible values are:
 
                 - `"no"`: No evaluation is done during training.
@@ -263,7 +263,7 @@ class TrainingArguments:
             requires more memory).
         eval_delay (`float`, *optional*):
             Number of epochs or steps to wait for before the first evaluation can be performed, depending on the
-            evaluation_strategy.
+            eval_strategy.
         learning_rate (`float`, *optional*, defaults to 5e-5):
             The initial learning rate for [`AdamW`] optimizer.
         weight_decay (`float`, *optional*, defaults to 0):
@@ -406,7 +406,7 @@ class TrainingArguments:
             Whether to drop the last incomplete batch (if the length of the dataset is not divisible by the batch size)
             or not.
         eval_steps (`int` or `float`, *optional*):
-            Number of update steps between two evaluations if `evaluation_strategy="steps"`. Will default to the same
+            Number of update steps between two evaluations if `eval_strategy="steps"`. Will default to the same
             value as `logging_steps` if not set. Should be an integer or a float in range `[0,1)`. If smaller than 1,
             will be interpreted as ratio of total training steps.
         dataloader_num_workers (`int`, *optional*, defaults to 0):
@@ -440,7 +440,7 @@ class TrainingArguments:
 
             <Tip>
 
-            When set to `True`, the parameters `save_strategy` needs to be the same as `evaluation_strategy`, and in
+            When set to `True`, the parameters `save_strategy` needs to be the same as `eval_strategy`, and in
             the case it is "steps", `save_steps` must be a round multiple of `eval_steps`.
 
             </Tip>
@@ -767,7 +767,7 @@ class TrainingArguments:
     do_train: bool = field(default=False, metadata={"help": "Whether to run training."})
     do_eval: bool = field(default=False, metadata={"help": "Whether to run eval on the dev set."})
     do_predict: bool = field(default=False, metadata={"help": "Whether to run predictions on the test set."})
-    evaluation_strategy: Union[IntervalStrategy, str] = field(
+    eval_strategy: Union[IntervalStrategy, str] = field(
         default="no",
         metadata={"help": "The evaluation strategy to use."},
     )
@@ -816,7 +816,7 @@ class TrainingArguments:
         metadata={
             "help": (
                 "Number of epochs or steps to wait for before the first evaluation can be performed, depending on the"
-                " evaluation_strategy."
+                " eval_strategy."
             )
         },
     )
@@ -1306,6 +1306,10 @@ class TrainingArguments:
             "choices": ["auto", "apex", "cpu_amp"],
         },
     )
+    evaluation_strategy: Union[IntervalStrategy, str] = field(
+        default=None,
+        metadata={"help": "Deprecated. Use `eval_strategy` instead"},
+    )
     push_to_hub_model_id: Optional[str] = field(
         default=None, metadata={"help": "The name of the repository to which push the `Trainer`."}
     )
@@ -1441,14 +1445,21 @@ class TrainingArguments:
         if self.disable_tqdm is None:
             self.disable_tqdm = logger.getEffectiveLevel() > logging.WARN
 
-        if isinstance(self.evaluation_strategy, EvaluationStrategy):
+        if self.evaluation_strategy is not None:
+            warnings.warn(
+                "`evaluation_strategy` is deprecated and will be removed in version 4.46 of 🤗 Transformers. Use `eval_strategy` instead",
+                FutureWarning,
+            )
+            self.eval_strategy = self.evaluation_strategy
+
+        if isinstance(self.eval_strategy, EvaluationStrategy):
             warnings.warn(
-                "using `EvaluationStrategy` for `evaluation_strategy` is deprecated and will be removed in version 5"
+                "using `EvaluationStrategy` for `eval_strategy` is deprecated and will be removed in version 5"
                 " of 🤗 Transformers. Use `IntervalStrategy` instead",
                 FutureWarning,
             )
             # Go back to the underlying string or we won't be able to instantiate `IntervalStrategy` on it.
-            self.evaluation_strategy = self.evaluation_strategy.value
+            self.eval_strategy = self.eval_strategy.value
         if self.no_cuda:
             warnings.warn(
                 "using `no_cuda` is deprecated and will be removed in version 5.0 of 🤗 Transformers. "
@@ -1457,23 +1468,23 @@ class TrainingArguments:
             )
             self.use_cpu = self.no_cuda
 
-        self.evaluation_strategy = IntervalStrategy(self.evaluation_strategy)
+        self.eval_strategy = IntervalStrategy(self.eval_strategy)
         self.logging_strategy = IntervalStrategy(self.logging_strategy)
         self.save_strategy = IntervalStrategy(self.save_strategy)
         self.hub_strategy = HubStrategy(self.hub_strategy)
 
         self.lr_scheduler_type = SchedulerType(self.lr_scheduler_type)
-        if self.do_eval is False and self.evaluation_strategy != IntervalStrategy.NO:
+        if self.do_eval is False and self.eval_strategy != IntervalStrategy.NO:
             self.do_eval = True
 
         # eval_steps has to be defined and non-zero, fallbacks to logging_steps if the latter is non-zero
-        if self.evaluation_strategy == IntervalStrategy.STEPS and (self.eval_steps is None or self.eval_steps == 0):
+        if self.eval_strategy == IntervalStrategy.STEPS and (self.eval_steps is None or self.eval_steps == 0):
             if self.logging_steps > 0:
                 logger.info(f"using `logging_steps` to initialize `eval_steps` to {self.logging_steps}")
                 self.eval_steps = self.logging_steps
             else:
                 raise ValueError(
-                    f"evaluation strategy {self.evaluation_strategy} requires either non-zero --eval_steps or"
+                    f"evaluation strategy {self.eval_strategy} requires either non-zero --eval_steps or"
                     " --logging_steps"
                 )
 
@@ -1485,7 +1496,7 @@ class TrainingArguments:
             if self.logging_steps != int(self.logging_steps):
                 raise ValueError(f"--logging_steps must be an integer if bigger than 1: {self.logging_steps}")
             self.logging_steps = int(self.logging_steps)
-        if self.evaluation_strategy == IntervalStrategy.STEPS and self.eval_steps > 1:
+        if self.eval_strategy == IntervalStrategy.STEPS and self.eval_steps > 1:
             if self.eval_steps != int(self.eval_steps):
                 raise ValueError(f"--eval_steps must be an integer if bigger than 1: {self.eval_steps}")
             self.eval_steps = int(self.eval_steps)
@@ -1496,12 +1507,12 @@ class TrainingArguments:
 
         # Sanity checks for load_best_model_at_end: we require save and eval strategies to be compatible.
         if self.load_best_model_at_end:
-            if self.evaluation_strategy != self.save_strategy:
+            if self.eval_strategy != self.save_strategy:
                 raise ValueError(
                     "--load_best_model_at_end requires the save and eval strategy to match, but found\n- Evaluation "
-                    f"strategy: {self.evaluation_strategy}\n- Save strategy: {self.save_strategy}"
+                    f"strategy: {self.eval_strategy}\n- Save strategy: {self.save_strategy}"
                 )
-            if self.evaluation_strategy == IntervalStrategy.STEPS and self.save_steps % self.eval_steps != 0:
+            if self.eval_strategy == IntervalStrategy.STEPS and self.save_steps % self.eval_steps != 0:
                 if self.eval_steps < 1 or self.save_steps < 1:
                     if not (self.eval_steps < 1 and self.save_steps < 1):
                         raise ValueError(
@@ -1579,7 +1590,7 @@ class TrainingArguments:
                 raise ValueError(" `--half_precision_backend apex`: GPU bf16 is not supported by apex.")
 
         if self.lr_scheduler_type == SchedulerType.REDUCE_ON_PLATEAU:
-            if self.evaluation_strategy == IntervalStrategy.NO:
+            if self.eval_strategy == IntervalStrategy.NO:
                 raise ValueError("lr_scheduler_type reduce_lr_on_plateau requires an eval strategy")
             if not is_torch_available():
                 raise ValueError("lr_scheduler_type reduce_lr_on_plateau requires torch>=0.2.0")
@@ -2443,7 +2454,7 @@ class TrainingArguments:
                 but requires more memory).
             delay (`float`, *optional*):
                 Number of epochs or steps to wait for before the first evaluation can be performed, depending on the
-                evaluation_strategy.
+                eval_strategy.
             loss_only (`bool`, *optional*, defaults to `False`):
                 Ignores all outputs except the loss.
             jit_mode (`bool`, *optional*):
@@ -2460,10 +2471,10 @@ class TrainingArguments:
         100
         ```
         """
-        self.evaluation_strategy = IntervalStrategy(strategy)
-        if self.evaluation_strategy == IntervalStrategy.STEPS and steps == 0:
+        self.eval_strategy = IntervalStrategy(strategy)
+        if self.eval_strategy == IntervalStrategy.STEPS and steps == 0:
             raise ValueError("Setting `strategy` as 'steps' requires a positive value for `steps`.")
-        self.do_eval = self.evaluation_strategy != IntervalStrategy.NO
+        self.do_eval = self.eval_strategy != IntervalStrategy.NO
         self.eval_steps = steps
         self.per_device_eval_batch_size = batch_size
         self.eval_accumulation_steps = accumulation_steps

src/transformers/training_args_tf.py

@@ -49,7 +49,7 @@ class TFTrainingArguments(TrainingArguments):
             by your training/evaluation scripts instead. See the [example
             scripts](https://github.com/huggingface/transformers/tree/main/examples) for more details.
         do_eval (`bool`, *optional*):
-            Whether to run evaluation on the validation set or not. Will be set to `True` if `evaluation_strategy` is
+            Whether to run evaluation on the validation set or not. Will be set to `True` if `eval_strategy` is
             different from `"no"`. This argument is not directly used by [`Trainer`], it's intended to be used by your
             training/evaluation scripts instead. See the [example
             scripts](https://github.com/huggingface/transformers/tree/main/examples) for more details.
@@ -57,7 +57,7 @@ class TFTrainingArguments(TrainingArguments):
             Whether to run predictions on the test set or not. This argument is not directly used by [`Trainer`], it's
             intended to be used by your training/evaluation scripts instead. See the [example
             scripts](https://github.com/huggingface/transformers/tree/main/examples) for more details.
-        evaluation_strategy (`str` or [`~trainer_utils.IntervalStrategy`], *optional*, defaults to `"no"`):
+        eval_strategy (`str` or [`~trainer_utils.IntervalStrategy`], *optional*, defaults to `"no"`):
             The evaluation strategy to adopt during training. Possible values are:
 
                 - `"no"`: No evaluation is done during training.

src/transformers/utils/notebook.py

@@ -292,11 +292,11 @@ class NotebookProgressCallback(TrainerCallback):
         self._force_next_update = False
 
     def on_train_begin(self, args, state, control, **kwargs):
-        self.first_column = "Epoch" if args.evaluation_strategy == IntervalStrategy.EPOCH else "Step"
+        self.first_column = "Epoch" if args.eval_strategy == IntervalStrategy.EPOCH else "Step"
         self.training_loss = 0
         self.last_log = 0
         column_names = [self.first_column] + ["Training Loss"]
-        if args.evaluation_strategy != IntervalStrategy.NO:
+        if args.eval_strategy != IntervalStrategy.NO:
             column_names.append("Validation Loss")
         self.training_tracker = NotebookTrainingTracker(state.max_steps, column_names)
 
@@ -328,7 +328,7 @@ class NotebookProgressCallback(TrainerCallback):
 
     def on_log(self, args, state, control, logs=None, **kwargs):
         # Only for when there is no evaluation
-        if args.evaluation_strategy == IntervalStrategy.NO and "loss" in logs:
+        if args.eval_strategy == IntervalStrategy.NO and "loss" in logs:
             values = {"Training Loss": logs["loss"]}
             # First column is necessarily Step sine we're not in epoch eval strategy
             values["Step"] = state.global_step

tests/deepspeed/test_deepspeed.py

@@ -959,7 +959,7 @@ class TrainerIntegrationDeepSpeed(TrainerIntegrationDeepSpeedWithCustomConfig, T
                 "do_train": True,
                 "do_eval": True,
                 "optim": "adafactor",
-                "evaluation_strategy": "steps",
+                "eval_strategy": "steps",
                 "eval_steps": 1,
                 "save_strategy": "steps",
                 "save_steps": 1,

tests/extended/test_trainer_ext.py

@@ -308,7 +308,7 @@ class TestTrainerExt(TestCasePlus):
             --per_device_eval_batch_size 4
             --max_eval_samples 8
             --val_max_target_length {max_len}
-            --evaluation_strategy steps
+            --eval_strategy steps
             --eval_steps {str(eval_steps)}
         """.split()
 

tests/fsdp/test_fsdp.py

@@ -308,6 +308,6 @@ class TrainerIntegrationFSDP(TestCasePlus, TrainerIntegrationCommon):
             --logging_steps {logging_steps}
             --save_strategy epoch
             --do_eval
-            --evaluation_strategy epoch
+            --eval_strategy epoch
             --report_to none
         """

tests/trainer/test_trainer.py

@@ -740,7 +740,7 @@ class TrainerIntegrationPrerunTest(TestCasePlus, TrainerIntegrationCommon):
         eval_dataset = RegressionDataset(length=64)
         args = TrainingArguments(
             "./regression",
-            evaluation_strategy="epoch",
+            eval_strategy="epoch",
             metric_for_best_model="eval_loss",
         )
         model = RegressionModel()
@@ -772,7 +772,7 @@ class TrainerIntegrationPrerunTest(TestCasePlus, TrainerIntegrationCommon):
         args = TrainingArguments(
             "./regression",
             lr_scheduler_type="reduce_lr_on_plateau",
-            evaluation_strategy="epoch",
+            eval_strategy="epoch",
             metric_for_best_model="eval_loss",
             num_train_epochs=10,
             learning_rate=0.2,
@@ -2210,7 +2210,7 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
                 output_dir=tmpdir,
                 learning_rate=0.1,
                 eval_steps=5,
-                evaluation_strategy="steps",
+                eval_strategy="steps",
                 save_steps=5,
                 load_best_model_at_end=True,
             )
@@ -2226,7 +2226,7 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
                 output_dir=tmpdir,
                 learning_rate=0.1,
                 eval_steps=5,
-                evaluation_strategy="steps",
+                eval_strategy="steps",
                 save_steps=5,
                 load_best_model_at_end=True,
                 metric_for_best_model="accuracy",
@@ -2243,7 +2243,7 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
                 b=2.5,
                 output_dir=tmpdir,
                 learning_rate=0.1,
-                evaluation_strategy="epoch",
+                eval_strategy="epoch",
                 save_strategy="epoch",
                 load_best_model_at_end=True,
                 metric_for_best_model="accuracy",
@@ -2262,7 +2262,7 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
                 output_dir=tmpdir,
                 learning_rate=0.1,
                 eval_steps=5,
-                evaluation_strategy="steps",
+                eval_strategy="steps",
                 save_steps=5,
                 load_best_model_at_end=True,
                 pretrained=False,
@@ -2283,7 +2283,7 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
                     output_dir=tmpdir,
                     learning_rate=0.1,
                     eval_steps=5,
-                    evaluation_strategy="steps",
+                    eval_strategy="steps",
                     save_steps=5,
                     load_best_model_at_end=True,
                     save_safetensors=save_safetensors,
@@ -2437,7 +2437,7 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
                 gradient_accumulation_steps=1,
                 per_device_train_batch_size=16,
                 load_best_model_at_end=True,
-                evaluation_strategy=IntervalStrategy.EPOCH,
+                eval_strategy=IntervalStrategy.EPOCH,
                 save_strategy=IntervalStrategy.EPOCH,
                 compute_metrics=AlmostAccuracy(),
                 metric_for_best_model="accuracy",
@@ -2453,7 +2453,7 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
                 num_train_epochs=20,
                 gradient_accumulation_steps=1,
                 per_device_train_batch_size=16,
-                evaluation_strategy=IntervalStrategy.EPOCH,
+                eval_strategy=IntervalStrategy.EPOCH,
                 compute_metrics=AlmostAccuracy(),
                 metric_for_best_model="accuracy",
             )
@@ -2497,7 +2497,7 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
 
             # With best model at end
             trainer = get_regression_trainer(
-                output_dir=tmp_dir, evaluation_strategy="steps", load_best_model_at_end=True, save_total_limit=2
+                output_dir=tmp_dir, eval_strategy="steps", load_best_model_at_end=True, save_total_limit=2
             )
             trainer.state.best_model_checkpoint = os.path.join(tmp_dir, "checkpoint-5")
             self.check_checkpoint_deletion(trainer, tmp_dir, [5, 25])
@@ -2505,7 +2505,7 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
             # Edge case: we don't always honor save_total_limit=1 if load_best_model_at_end=True to be able to resume
             # from checkpoint
             trainer = get_regression_trainer(
-                output_dir=tmp_dir, evaluation_strategy="steps", load_best_model_at_end=True, save_total_limit=1
+                output_dir=tmp_dir, eval_strategy="steps", load_best_model_at_end=True, save_total_limit=1
             )
             trainer.state.best_model_checkpoint = os.path.join(tmp_dir, "checkpoint-25")
             self.check_checkpoint_deletion(trainer, tmp_dir, [25])
@@ -3341,7 +3341,7 @@ class TrainerHyperParameterOptunaIntegrationTest(unittest.TestCase):
                 output_dir=tmp_dir,
                 learning_rate=0.1,
                 logging_steps=1,
-                evaluation_strategy=IntervalStrategy.EPOCH,
+                eval_strategy=IntervalStrategy.EPOCH,
                 save_strategy=IntervalStrategy.EPOCH,
                 num_train_epochs=4,
                 disable_tqdm=True,
@@ -3390,7 +3390,7 @@ class TrainerHyperParameterMultiObjectOptunaIntegrationTest(unittest.TestCase):
                 output_dir=tmp_dir,
                 learning_rate=0.1,
                 logging_steps=1,
-                evaluation_strategy=IntervalStrategy.EPOCH,
+                eval_strategy=IntervalStrategy.EPOCH,
                 save_strategy=IntervalStrategy.EPOCH,
                 num_train_epochs=10,
                 disable_tqdm=True,
@@ -3448,7 +3448,7 @@ class TrainerHyperParameterRayIntegrationTest(unittest.TestCase):
                 output_dir=tmp_dir,
                 learning_rate=0.1,
                 logging_steps=1,
-                evaluation_strategy=IntervalStrategy.EPOCH,
+                eval_strategy=IntervalStrategy.EPOCH,
                 save_strategy=IntervalStrategy.EPOCH,
                 num_train_epochs=4,
                 disable_tqdm=True,
@@ -3511,7 +3511,7 @@ class TrainerHyperParameterSigOptIntegrationTest(unittest.TestCase):
                 output_dir=tmp_dir,
                 learning_rate=0.1,
                 logging_steps=1,
-                evaluation_strategy=IntervalStrategy.EPOCH,
+                eval_strategy=IntervalStrategy.EPOCH,
                 save_strategy=IntervalStrategy.EPOCH,
                 num_train_epochs=4,
                 disable_tqdm=True,
@@ -3931,7 +3931,7 @@ class TrainerHyperParameterWandbIntegrationTest(unittest.TestCase):
                 output_dir=tmp_dir,
                 learning_rate=0.1,
                 logging_steps=1,
-                evaluation_strategy=IntervalStrategy.EPOCH,
+                eval_strategy=IntervalStrategy.EPOCH,
                 save_strategy=IntervalStrategy.EPOCH,
                 num_train_epochs=4,
                 disable_tqdm=True,

tests/trainer/test_trainer_callback.py

@@ -133,12 +133,12 @@ class TrainerCallbackTest(unittest.TestCase):
                 expected_events += ["on_step_begin", "on_step_end"]
                 if step % trainer.args.logging_steps == 0:
                     expected_events.append("on_log")
-                if trainer.args.evaluation_strategy == IntervalStrategy.STEPS and step % trainer.args.eval_steps == 0:
+                if trainer.args.eval_strategy == IntervalStrategy.STEPS and step % trainer.args.eval_steps == 0:
                     expected_events += evaluation_events.copy()
                 if step % trainer.args.save_steps == 0:
                     expected_events.append("on_save")
             expected_events.append("on_epoch_end")
-            if trainer.args.evaluation_strategy == IntervalStrategy.EPOCH:
+            if trainer.args.eval_strategy == IntervalStrategy.EPOCH:
                 expected_events += evaluation_events.copy()
         expected_events += ["on_log", "on_train_end"]
         return expected_events
@@ -215,12 +215,12 @@ class TrainerCallbackTest(unittest.TestCase):
         events = trainer.callback_handler.callbacks[-2].events
         self.assertEqual(events, self.get_expected_events(trainer))
 
-        trainer = self.get_trainer(callbacks=[MyTestTrainerCallback], eval_steps=5, evaluation_strategy="steps")
+        trainer = self.get_trainer(callbacks=[MyTestTrainerCallback], eval_steps=5, eval_strategy="steps")
         trainer.train()
         events = trainer.callback_handler.callbacks[-2].events
         self.assertEqual(events, self.get_expected_events(trainer))
 
-        trainer = self.get_trainer(callbacks=[MyTestTrainerCallback], evaluation_strategy="epoch")
+        trainer = self.get_trainer(callbacks=[MyTestTrainerCallback], eval_strategy="epoch")
         trainer.train()
         events = trainer.callback_handler.callbacks[-2].events
         self.assertEqual(events, self.get_expected_events(trainer))
@@ -231,7 +231,7 @@ class TrainerCallbackTest(unittest.TestCase):
             logging_steps=3,
             save_steps=10,
             eval_steps=5,
-            evaluation_strategy="steps",
+            eval_strategy="steps",
         )
         trainer.train()
         events = trainer.callback_handler.callbacks[-2].events

tests/trainer/test_trainer_seq2seq.py

@@ -113,7 +113,7 @@ class Seq2seqTrainerTester(TestCasePlus):
             per_device_train_batch_size=batch_size,
             per_device_eval_batch_size=batch_size,
             predict_with_generate=True,
-            evaluation_strategy="steps",
+            eval_strategy="steps",
             do_train=True,
             do_eval=True,
             warmup_steps=0,