Set cache_dir for evaluate.load() in example scripts (#28422)

While using `run_clm.py`,[^1] I noticed that some files were being added
to my global cache, not the local cache. I set the `cache_dir` parameter
for the one call to `evaluate.load()`, which partially solved the
problem. I figured that while I was fixing the one script upstream, I
might as well fix the problem in all other example scripts that I could.

There are still some files being added to my global cache, but this
appears to be a bug in `evaluate` itself. This commit at least moves
some of the files into the local cache, which is better than before.

To create this PR, I made the following regex-based transformation:
`evaluate\.load\((.*?)\)` -> `evaluate\.load\($1,
cache_dir=model_args.cache_dir\)`. After using that, I manually fixed
all modified files with `ruff` serving as useful guidance. During the
process, I removed one existing usage of the `cache_dir` parameter in a
script that did not have a corresponding `--cache-dir` argument
declared.

[^1]: I specifically used `pytorch/language-modeling/run_clm.py` from
v4.34.1 of the library. For the original code, see the following URL:
https://github.com/huggingface/transformers/tree/acc394c4f5e1283c19783581790b3dc3105a3697/examples/pytorch/language-modeling/run_clm.py.

examples/flax/image-captioning/run_image_captioning_flax.py

@@ -853,7 +853,7 @@ def blockwise_data_loader(
                 yield batch
 
     # Metric
-    metric = evaluate.load("rouge")
+    metric = evaluate.load("rouge", cache_dir=model_args.cache_dir)
 
     def postprocess_text(preds, labels):
         preds = [pred.strip() for pred in preds]

examples/flax/question-answering/run_qa.py

@@ -807,7 +807,9 @@ def post_processing_function(examples, features, predictions, stage="eval"):
         references = [{"id": ex["id"], "answers": ex[answer_column_name]} for ex in examples]
         return EvalPrediction(predictions=formatted_predictions, label_ids=references)
 
-    metric = evaluate.load("squad_v2" if data_args.version_2_with_negative else "squad")
+    metric = evaluate.load(
+        "squad_v2" if data_args.version_2_with_negative else "squad", cache_dir=model_args.cache_dir
+    )
 
     def compute_metrics(p: EvalPrediction):
         return metric.compute(predictions=p.predictions, references=p.label_ids)

examples/flax/speech-recognition/run_flax_speech_recognition_seq2seq.py

@@ -577,7 +577,7 @@ def is_audio_in_length_range(length):
         return
 
     # 8. Load Metric
-    metric = evaluate.load("wer")
+    metric = evaluate.load("wer", cache_dir=model_args.cache_dir)
 
     def compute_metrics(preds, labels):
         # replace padded labels by the padding token

examples/flax/summarization/run_summarization_flax.py

@@ -710,7 +710,7 @@ def preprocess_function(examples):
         )
 
     # Metric
-    metric = evaluate.load("rouge")
+    metric = evaluate.load("rouge", cache_dir=model_args.cache_dir)
 
     def postprocess_text(preds, labels):
         preds = [pred.strip() for pred in preds]

examples/flax/text-classification/run_flax_glue.py

@@ -599,9 +599,9 @@ def eval_step(state, batch):
     p_eval_step = jax.pmap(eval_step, axis_name="batch")
 
     if data_args.task_name is not None:
-        metric = evaluate.load("glue", data_args.task_name)
+        metric = evaluate.load("glue", data_args.task_name, cache_dir=model_args.cache_dir)
     else:
-        metric = evaluate.load("accuracy")
+        metric = evaluate.load("accuracy", cache_dir=model_args.cache_dir)
 
     logger.info(f"===== Starting training ({num_epochs} epochs) =====")
     train_time = 0

examples/flax/token-classification/run_flax_ner.py

@@ -676,7 +676,7 @@ def eval_step(state, batch):
 
     p_eval_step = jax.pmap(eval_step, axis_name="batch")
 
-    metric = evaluate.load("seqeval")
+    metric = evaluate.load("seqeval", cache_dir=model_args.cache_dir)
 
     def get_labels(y_pred, y_true):
         # Transform predictions and references tensos to numpy arrays

examples/pytorch/audio-classification/run_audio_classification.py

@@ -349,7 +349,7 @@ def val_transforms(batch):
         id2label[str(i)] = label
 
     # Load the accuracy metric from the datasets package
-    metric = evaluate.load("accuracy")
+    metric = evaluate.load("accuracy", cache_dir=model_args.cache_dir)
 
     # Define our compute_metrics function. It takes an `EvalPrediction` object (a namedtuple with
     # `predictions` and `label_ids` fields) and has to return a dictionary string to float.

examples/pytorch/image-classification/run_image_classification.py

@@ -287,7 +287,7 @@ def main():
         id2label[str(i)] = label
 
     # Load the accuracy metric from the datasets package
-    metric = evaluate.load("accuracy")
+    metric = evaluate.load("accuracy", cache_dir=model_args.cache_dir)
 
     # Define our compute_metrics function. It takes an `EvalPrediction` object (a namedtuple with a
     # predictions and label_ids field) and has to return a dictionary string to float.

examples/pytorch/image-classification/run_image_classification_no_trainer.py

@@ -282,7 +282,6 @@ def main():
         dataset = load_dataset(
             "imagefolder",
             data_files=data_files,
-            cache_dir=args.cache_dir,
             task="image-classification",
         )
         # See more about loading custom images at

examples/pytorch/language-modeling/run_clm.py

@@ -583,7 +583,7 @@ def preprocess_logits_for_metrics(logits, labels):
                 logits = logits[0]
             return logits.argmax(dim=-1)
 
-        metric = evaluate.load("accuracy")
+        metric = evaluate.load("accuracy", cache_dir=model_args.cache_dir)
 
         def compute_metrics(eval_preds):
             preds, labels = eval_preds

examples/pytorch/language-modeling/run_mlm.py

@@ -590,7 +590,7 @@ def preprocess_logits_for_metrics(logits, labels):
                 logits = logits[0]
             return logits.argmax(dim=-1)
 
-        metric = evaluate.load("accuracy")
+        metric = evaluate.load("accuracy", cache_dir=model_args.cache_dir)
 
         def compute_metrics(eval_preds):
             preds, labels = eval_preds

examples/pytorch/question-answering/run_qa.py

@@ -627,7 +627,9 @@ def post_processing_function(examples, features, predictions, stage="eval"):
         references = [{"id": str(ex["id"]), "answers": ex[answer_column_name]} for ex in examples]
         return EvalPrediction(predictions=formatted_predictions, label_ids=references)
 
-    metric = evaluate.load("squad_v2" if data_args.version_2_with_negative else "squad")
+    metric = evaluate.load(
+        "squad_v2" if data_args.version_2_with_negative else "squad", cache_dir=model_args.cache_dir
+    )
 
     def compute_metrics(p: EvalPrediction):
         return metric.compute(predictions=p.predictions, references=p.label_ids)

examples/pytorch/question-answering/run_qa_beam_search.py

@@ -647,7 +647,9 @@ def post_processing_function(examples, features, predictions, stage="eval"):
         references = [{"id": ex["id"], "answers": ex[answer_column_name]} for ex in examples]
         return EvalPrediction(predictions=formatted_predictions, label_ids=references)
 
-    metric = evaluate.load("squad_v2" if data_args.version_2_with_negative else "squad")
+    metric = evaluate.load(
+        "squad_v2" if data_args.version_2_with_negative else "squad", cache_dir=model_args.cache_dir
+    )
 
     def compute_metrics(p: EvalPrediction):
         return metric.compute(predictions=p.predictions, references=p.label_ids)

examples/pytorch/question-answering/run_seq2seq_qa.py

@@ -631,7 +631,9 @@ def preprocess_validation_function(examples):
         pad_to_multiple_of=8 if training_args.fp16 else None,
     )
 
-    metric = evaluate.load("squad_v2" if data_args.version_2_with_negative else "squad")
+    metric = evaluate.load(
+        "squad_v2" if data_args.version_2_with_negative else "squad", cache_dir=model_args.cache_dir
+    )
 
     def compute_metrics(p: EvalPrediction):
         return metric.compute(predictions=p.predictions, references=p.label_ids)

examples/pytorch/semantic-segmentation/run_semantic_segmentation.py

@@ -366,7 +366,7 @@ def main():
     label2id = {v: str(k) for k, v in id2label.items()}
 
     # Load the mean IoU metric from the datasets package
-    metric = evaluate.load("mean_iou")
+    metric = evaluate.load("mean_iou", cache_dir=model_args.cache_dir)
 
     # Define our compute_metrics function. It takes an `EvalPrediction` object (a namedtuple with a
     # predictions and label_ids field) and has to return a dictionary string to float.

examples/pytorch/semantic-segmentation/run_semantic_segmentation_no_trainer.py

@@ -530,7 +530,7 @@ def preprocess_val(example_batch):
     args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
 
     # Instantiate metric
-    metric = evaluate.load("mean_iou")
+    metric = evaluate.load("mean_iou", cache_dir=args.cache_dir)
 
     # We need to initialize the trackers we use, and also store our configuration.
     # The trackers initializes automatically on the main process.

examples/pytorch/speech-recognition/run_speech_recognition_ctc.py

@@ -680,7 +680,7 @@ def is_audio_in_length_range(length):
     # instantiate a data collator and the trainer
 
     # Define evaluation metrics during training, *i.e.* word error rate, character error rate
-    eval_metrics = {metric: evaluate.load(metric) for metric in data_args.eval_metrics}
+    eval_metrics = {metric: evaluate.load(metric, cache_dir=model_args.cache_dir) for metric in data_args.eval_metrics}
 
     # for large datasets it is advised to run the preprocessing on a
     # single machine first with ``args.preprocessing_only`` since there will mostly likely

examples/pytorch/speech-recognition/run_speech_recognition_ctc_adapter.py

@@ -702,7 +702,7 @@ def is_audio_in_length_range(length):
     # instantiate a data collator and the trainer
 
     # Define evaluation metrics during training, *i.e.* word error rate, character error rate
-    eval_metrics = {metric: evaluate.load(metric) for metric in data_args.eval_metrics}
+    eval_metrics = {metric: evaluate.load(metric, cache_dir=model_args.cache_dir) for metric in data_args.eval_metrics}
 
     # for large datasets it is advised to run the preprocessing on a
     # single machine first with ``args.preprocessing_only`` since there will mostly likely

examples/pytorch/speech-recognition/run_speech_recognition_seq2seq.py

@@ -520,7 +520,7 @@ def is_audio_in_length_range(length):
         return
 
     # 8. Load Metric
-    metric = evaluate.load("wer")
+    metric = evaluate.load("wer", cache_dir=model_args.cache_dir)
 
     def compute_metrics(pred):
         pred_ids = pred.predictions

examples/pytorch/summarization/run_summarization.py

@@ -645,7 +645,7 @@ def preprocess_function(examples):
     )
 
     # Metric
-    metric = evaluate.load("rouge")
+    metric = evaluate.load("rouge", cache_dir=model_args.cache_dir)
 
     def postprocess_text(preds, labels):
         preds = [pred.strip() for pred in preds]

examples/pytorch/text-classification/run_classification.py

@@ -633,23 +633,23 @@ def preprocess_function(examples):
 
     if data_args.metric_name is not None:
         metric = (
-            evaluate.load(data_args.metric_name, config_name="multilabel")
+            evaluate.load(data_args.metric_name, config_name="multilabel", cache_dir=model_args.cache_dir)
             if is_multi_label
-            else evaluate.load(data_args.metric_name)
+            else evaluate.load(data_args.metric_name, cache_dir=model_args.cache_dir)
         )
         logger.info(f"Using metric {data_args.metric_name} for evaluation.")
     else:
         if is_regression:
-            metric = evaluate.load("mse")
+            metric = evaluate.load("mse", cache_dir=model_args.cache_dir)
             logger.info("Using mean squared error (mse) as regression score, you can use --metric_name to overwrite.")
         else:
             if is_multi_label:
-                metric = evaluate.load("f1", config_name="multilabel")
+                metric = evaluate.load("f1", config_name="multilabel", cache_dir=model_args.cache_dir)
                 logger.info(
                     "Using multilabel F1 for multi-label classification task, you can use --metric_name to overwrite."
                 )
             else:
-                metric = evaluate.load("accuracy")
+                metric = evaluate.load("accuracy", cache_dir=model_args.cache_dir)
                 logger.info("Using accuracy as classification score, you can use --metric_name to overwrite.")
 
     def compute_metrics(p: EvalPrediction):

examples/pytorch/text-classification/run_glue.py

@@ -514,11 +514,11 @@ def preprocess_function(examples):
 
     # Get the metric function
     if data_args.task_name is not None:
-        metric = evaluate.load("glue", data_args.task_name)
+        metric = evaluate.load("glue", data_args.task_name, cache_dir=model_args.cache_dir)
     elif is_regression:
-        metric = evaluate.load("mse")
+        metric = evaluate.load("mse", cache_dir=model_args.cache_dir)
     else:
-        metric = evaluate.load("accuracy")
+        metric = evaluate.load("accuracy", cache_dir=model_args.cache_dir)
 
     # You can define your custom compute_metrics function. It takes an `EvalPrediction` object (a namedtuple with a
     # predictions and label_ids field) and has to return a dictionary string to float.

examples/pytorch/text-classification/run_xnli.py

@@ -385,7 +385,7 @@ def preprocess_function(examples):
             )
 
     # Get the metric function
-    metric = evaluate.load("xnli")
+    metric = evaluate.load("xnli", cache_dir=model_args.cache_dir)
 
     # You can define your custom compute_metrics function. It takes an `EvalPrediction` object (a namedtuple with a
     # predictions and label_ids field) and has to return a dictionary string to float.

examples/pytorch/token-classification/run_ner.py

@@ -539,7 +539,7 @@ def tokenize_and_align_labels(examples):
     data_collator = DataCollatorForTokenClassification(tokenizer, pad_to_multiple_of=8 if training_args.fp16 else None)
 
     # Metrics
-    metric = evaluate.load("seqeval")
+    metric = evaluate.load("seqeval", cache_dir=model_args.cache_dir)
 
     def compute_metrics(p):
         predictions, labels = p

examples/pytorch/translation/run_translation.py

@@ -564,7 +564,7 @@ def preprocess_function(examples):
         )
 
     # Metric
-    metric = evaluate.load("sacrebleu")
+    metric = evaluate.load("sacrebleu", cache_dir=model_args.cache_dir)
 
     def postprocess_text(preds, labels):
         preds = [pred.strip() for pred in preds]

examples/tensorflow/image-classification/run_image_classification.py

@@ -440,7 +440,7 @@ def val_transforms(example_batch):
     collate_fn = DefaultDataCollator(return_tensors="np")
 
     # Load the accuracy metric from the datasets package
-    metric = evaluate.load("accuracy")
+    metric = evaluate.load("accuracy", cache_dir=model_args.cache_dir)
 
     # Define our compute_metrics function. It takes an `EvalPrediction` object (a namedtuple with a
     # predictions and label_ids field) and has to return a dictionary string to float.

examples/tensorflow/question-answering/run_qa.py

@@ -631,7 +631,9 @@ def post_processing_function(examples, features, predictions, stage="eval"):
         references = [{"id": ex["id"], "answers": ex[answer_column_name]} for ex in examples]
         return EvalPrediction(predictions=formatted_predictions, label_ids=references)
 
-    metric = evaluate.load("squad_v2" if data_args.version_2_with_negative else "squad")
+    metric = evaluate.load(
+        "squad_v2" if data_args.version_2_with_negative else "squad", cache_dir=model_args.cache_dir
+    )
 
     def compute_metrics(p: EvalPrediction):
         return metric.compute(predictions=p.predictions, references=p.label_ids)

examples/tensorflow/summarization/run_summarization.py

@@ -627,7 +627,7 @@ def postprocess_text(preds, labels):
 
         # region Metric and KerasMetricCallback
         if training_args.do_eval:
-            metric = evaluate.load("rouge")
+            metric = evaluate.load("rouge", cache_dir=model_args.cache_dir)
 
             if data_args.val_max_target_length is None:
                 data_args.val_max_target_length = data_args.max_target_length

examples/tensorflow/text-classification/run_glue.py

@@ -379,7 +379,7 @@ def preprocess_function(examples):
     # endregion
 
     # region Metric function
-    metric = evaluate.load("glue", data_args.task_name)
+    metric = evaluate.load("glue", data_args.task_name, cache_dir=model_args.cache_dir)
 
     def compute_metrics(preds, label_ids):
         preds = preds["logits"]

examples/tensorflow/token-classification/run_ner.py

@@ -511,7 +511,7 @@ def tokenize_and_align_labels(examples):
         # endregion
 
         # Metrics
-        metric = evaluate.load("seqeval")
+        metric = evaluate.load("seqeval", cache_dir=model_args.cache_dir)
 
         def get_labels(y_pred, y_true):
             # Transform predictions and references tensos to numpy arrays

examples/tensorflow/translation/run_translation.py

@@ -589,7 +589,7 @@ def preprocess_function(examples):
 
         # region Metric and postprocessing
         if training_args.do_eval:
-            metric = evaluate.load("sacrebleu")
+            metric = evaluate.load("sacrebleu", cache_dir=model_args.cache_dir)
 
             if data_args.val_max_target_length is None:
                 data_args.val_max_target_length = data_args.max_target_length