Update tasdda.py

tasdda.py

@@ -1,342 +1,343 @@
-import csv
-import datasets
-from datasets import BuilderConfig, GeneratorBasedBuilder, DatasetInfo, SplitGenerator, Split
-from pathlib import Path
-
-_PROMPTS_PROSODIC_URLS = {
-    "dev": "prosodic/validation.csv",
-    "train": "prosodic/train.csv",
-}
-
-_PROMPTS_AUDIO_CORPUS_URLS = {
-    "dev": "audioCorpus/validation.csv",
-    "train": "audioCorpus/train.csv",
-}
-
-_PROMPTS_AUTOMATIC_URLS = {
-    "dev": "automatic/validation.csv",
-    "train": "automatic/train.csv",
-}
-
-_PROMPTS_TEST_URLS = {
-    "test": "test/test.csv",
-}
-
-_ARCHIVES_PROSODIC = {
-    "dev": "prosodic/audios.tar.gz",
-    "train": "prosodic/audios.tar.gz",
-}
-
-_ARCHIVES_AUDIO_CORPUS = {
-    "dev": "audioCorpus/audios.tar.gz",
-    "train": "audioCorpus/audios.tar.gz",
-}
-
-_ARCHIVES_AUTOMATIC = {
-    "train": "automatic/nurc_cm_automatic_segmented_audios.zip",
-    "dev": "automatic/nurc_cm_automatic_segmented_audios.zip",
-}
-
-_ARCHIVES_TEST = {
-    "test": "test/test.zip",
-}
-
-_PATH_TO_CLIPS = {
-    "dev": "",
-    "train": "",
-    "test": "",
-}
-
-
-class NurcSPConfig(BuilderConfig):
-    def __init__(self, prompts_type, **kwargs):
-        super().__init__(**kwargs)
-        self.prompts_type = prompts_type
-
-
-class NurcSPDataset(GeneratorBasedBuilder):
-    BUILDER_CONFIGS = [
-        NurcSPConfig(name="audioCorpus", description="Audio Corpus audio prompts", prompts_type="audioCorpus"),
-        NurcSPConfig(name="prosodic", description="Prosodic audio prompts", prompts_type="prosodic"),
-        NurcSPConfig(name="automatic", description="Automatic audio prompts", prompts_type="automatic"),
-        NurcSPConfig(name="test", description="Test audio prompts", prompts_type="test"),
-    ]
-
-    def _info(self):
-        if self.config.name == "prosodic":
-            return DatasetInfo(
-                features=datasets.Features(
-                    {
-                        "path": datasets.Value("string"),
-                        "name": datasets.Value("string"),
-                        "speaker": datasets.Value("string"),
-                        "start_time": datasets.Value("string"),
-                        "end_time": datasets.Value("string"),
-                        "normalized_text": datasets.Value("string"),
-                        "text": datasets.Value("string"),
-                        "duration": datasets.Value("string"),
-                        "type": datasets.Value("string"),
-                        "year": datasets.Value("string"),
-                        "gender": datasets.Value("string"),
-                        "age_range": datasets.Value("string"),
-                        "total_duration": datasets.Value("string"),
-                        "quality": datasets.Value("string"),
-                        "theme": datasets.Value("string"),
-                        "audio": datasets.Audio(sampling_rate=16_000, mono=True),
-                    }
-                )
-            )
-        elif self.config.name == "audioCorpus":
-            return DatasetInfo(
-                features=datasets.Features(
-                    {
-                        "audio_name": datasets.Value("string"),
-                        "file_path": datasets.Value("string"),
-                        "text": datasets.Value("string"),
-                        "start_time": datasets.Value("string"),
-                        "end_time": datasets.Value("string"),
-                        "duration": datasets.Value("string"),
-                        "quality": datasets.Value("string"),
-                        "speech_genre": datasets.Value("string"),
-                        "speech_style": datasets.Value("string"),
-                        "variety": datasets.Value("string"),
-                        "accent": datasets.Value("string"),
-                        "sex": datasets.Value("string"),
-                        "age_range": datasets.Value("string"),
-                        "num_speakers": datasets.Value("string"),
-                        "speaker_id": datasets.Value("string"),
-                        "audio": datasets.Audio(sampling_rate=16_000, mono=True),
-                    }
-                )
-            )
-        elif self.config.name == "automatic":
-            return DatasetInfo(
-                features=datasets.Features(
-                    {
-                        "path": datasets.Value("string"),
-                        "name": datasets.Value("string"),
-                        "speaker": datasets.Value("string"),
-                        "start_time": datasets.Value("string"),
-                        "end_time": datasets.Value("string"),
-                        "text": datasets.Value("string"),
-                        "duration": datasets.Value("string"),
-                        "audio": datasets.Audio(sampling_rate=16_000, mono=True),
-                    }
-                )
-            )
-        elif self.config.name == "test":
-            return DatasetInfo(
-                features=datasets.Features(
-                    {
-                        "path": datasets.Value("string"),
-                        "name": datasets.Value("string"),
-                        "speaker": datasets.Value("string"),  
-                        "start_time": datasets.Value("string"),
-                        "end_time": datasets.Value("string"),
-                        "text": datasets.Value("string"),
-                        "duration": datasets.Value("string"),
-                        "audio": datasets.Audio(sampling_rate=16_000, mono=True),                  
-                    }
-                )
-            )
-
-
-    def _split_generators(self, dl_manager):
-        if self.config.prompts_type == "prosodic":
-            prompts_urls = _PROMPTS_PROSODIC_URLS
-            archive_link = _ARCHIVES_PROSODIC
-        elif self.config.prompts_type == "audioCorpus":
-            prompts_urls = _PROMPTS_AUDIO_CORPUS_URLS
-            archive_link = _ARCHIVES_AUDIO_CORPUS
-        elif self.config.prompts_type == "automatic":
-            prompts_urls = _PROMPTS_AUTOMATIC_URLS
-            archive_link = _ARCHIVES_AUTOMATIC
-        elif self.config.prompts_type == "test":
-            prompts_urls = _PROMPTS_TEST_URLS
-            archive_link = _ARCHIVES_TEST
-        else:
-            return
-
-        prompts_path = dl_manager.download(prompts_urls)
-        archive = dl_manager.download(archive_link)
-
-        if self.config.prompts_type == "prosodic" or self.config.prompts_type == "audioCorpus":
-            return [
-                SplitGenerator(
-                    name=Split.VALIDATION,
-                    gen_kwargs={
-                        "prompts_path": prompts_path["dev"],
-                        "path_to_clips": _PATH_TO_CLIPS["dev"],
-                        "audio_files": dl_manager.iter_archive(archive["dev"]),
-                        "split_name": "validation"
-                    }
-                ),
-                SplitGenerator(
-                    name=Split.TRAIN,
-                    gen_kwargs={
-                        "prompts_path": prompts_path["train"],
-                        "path_to_clips": _PATH_TO_CLIPS["train"],
-                        "audio_files": dl_manager.iter_archive(archive["train"]),
-                        "split_name": "train"
-                    }
-                ),
-            ]
-        elif self.config.prompts_type == "automatic":
-            return [
-                SplitGenerator(
-                    name=Split.VALIDATION,
-                    gen_kwargs={
-                        "prompts_path": prompts_path["dev"],
-                        "path_to_clips": _PATH_TO_CLIPS["dev"],
-                        "audio_files": dl_manager.iter_archive(archive["dev"]),
-                        "split_name": "validation"
-                    }
-                ),
-                SplitGenerator(
-                    name=Split.TRAIN,
-                    gen_kwargs={
-                        "prompts_path": prompts_path["train"],
-                        "path_to_clips": _PATH_TO_CLIPS["train"],
-                        "audio_files": dl_manager.iter_archive(archive["train"]),
-                        "split_name": "train"
-                    }
-                ),
-            ]
-        elif self.config.prompts_type == "test":
-            return[
-                SplitGenerator(
-                    name=Split.TEST,
-                    gen_kwargs={
-                        "prompts_path": prompts_path["test"],
-                        "path_to_clips": _PATH_TO_CLIPS["test"],
-                        "audio_files": dl_manager.iter_archive(archive["test"]),
-                        "split_name": "test"
-                    }
-                ),
-            ]
-
-    def _generate_examples(self, prompts_path, path_to_clips, audio_files, split_name):
-        examples = {}
-        csv_paths = []
-
-        with open(prompts_path, "r", encoding="utf-8") as f:
-            if self.config.prompts_type == "test":
-                csv_reader = csv.DictReader(f, delimiter=";")  # Explicitly set delimiter
-            else:
-                csv_reader = csv.DictReader(f)
-                
-            if self.config.prompts_type == "prosodic":
-                for row in csv_reader:
-                    file_path = Path(row['path']).as_posix()
-                    examples[file_path] = {
-                        "path": row['path'],
-                        "name": row['name'],
-                        "speaker": row['speaker'],
-                        "start_time": row['start_time'],
-                        "end_time": row['end_time'],
-                        "normalized_text": row['normalized_text'],
-                        "text": row['text'],
-                        "duration": row['duration'],
-                        "type": row['type'],
-                        "year": row['year'],
-                        "gender": row['gender'],
-                        "age_range": row['age_range'],
-                        "total_duration": row['total_duration'],
-                        "quality": row['quality'],
-                        "theme": row['theme'],
-                    }
-                    csv_paths.append(file_path)
-            elif self.config.prompts_type == "audioCorpus":
-                for row in csv_reader:
-                    file_path = Path(row['file_path']).as_posix()
-                    examples[file_path] = {
-                        "audio_name": row['audio_name'],
-                        "file_path": row['file_path'],
-                        "text": row['text'],
-                        "start_time": row['start_time'],
-                        "end_time": row['end_time'],
-                        "duration": row['duration'],
-                        "quality": row['quality'],
-                        "speech_genre": row['speech_genre'],
-                        "speech_style": row['speech_style'],
-                        "variety": row['variety'],
-                        "accent": row['accent'],
-                        "sex": row['sex'],
-                        "age_range": row['age_range'],
-                        "num_speakers": row['num_speakers'],
-                        "speaker_id": row['speaker_id'],
-                    }
-                    csv_paths.append(file_path)
-            elif self.config.prompts_type == "automatic":
-                for row in csv_reader:
-                    file_path = Path(row['path']).as_posix()
-                    examples[file_path] = {
-                        "path": row['path'],
-                        "name": row['name'],
-                        "speaker": row['speaker'],
-                        "start_time": row['start_time'],
-                        "end_time": row['end_time'],
-                        "text": row['text'],
-                        "duration": row['duration'],
-                    }
-                    csv_paths.append(file_path)
-            elif self.config.prompts_type == "test":
-                for row in csv_reader:
-                    file_path = Path(row['path']).as_posix()
-                    examples[file_path] = {
-                        "path": row['path'],
-                        "name": row['name'],
-                        "speaker": row['speaker'],
-                        "start_time": row['start_time'],
-                        "end_time": row['end_time'],
-                        "text": row['text'],
-                        "duration": row['duration'],
-                    }
-                    csv_paths.append(file_path)
-
-        id_ = 0
-        for path, f in audio_files:
-            path = Path(path).as_posix()
-            if path.startswith(path_to_clips) and path in examples:
-                audio = {"path": path, "bytes": f.read()}
-                yield id_, {**examples[path], "audio": audio}
-                id_ += 1
-
-
-if __name__ == "__main__":
-    from datasets import DatasetDict, DatasetInfo
-    import json
-    from pathlib import Path
-
-    infos = {}
-
-    for cfg in ["prosodic", "audioCorpus", "automatic", "test"]:
-        # Load parquet files into a DatasetDict
-        parquet_files = list(Path(f"parquet/{cfg}").glob("*.parquet"))
-        ds = DatasetDict({
-            f.stem.split("_")[-1]:  # e.g., "train", "validation", "test"
-            DatasetDict.from_parquet(str(f)) for f in parquet_files
-        })
-
-        # Build info object
-        info = DatasetInfo.from_dataset_dict(ds)
-
-        # Store in dict
-        infos[cfg] = info.to_dict()
-
-    # Save to dataset_infos.json
-    with open("dataset_infos.json", "w", encoding="utf-8") as f:
-        json.dump(infos, f, indent=2, ensure_ascii=False)
-
-    '''
-    for cfg in ["prosodic", "audioCorpus", "automatic", "test"]:
-        builder = NurcSPDataset(config_name=cfg)
-        builder.download_and_prepare()
-        ds: DatasetDict = builder.as_dataset()
-
-        out_dir = Path(f"parquet/{cfg}")
-        out_dir.mkdir(parents=True, exist_ok=True)
-
-        for split, dataset in ds.items():
-            dataset.to_parquet(out_dir / f"{cfg}_{split}.parquet")
+import csv
+import datasets
+from datasets import BuilderConfig, GeneratorBasedBuilder, DatasetInfo, SplitGenerator, Split
+from pathlib import Path
+
+
+_PROMPTS_PROSODIC_URLS = {
+    "dev": "prosodic/validation.csv",
+    "train": "prosodic/train.csv",
+}
+
+_PROMPTS_AUDIO_CORPUS_URLS = {
+    "dev": "audioCorpus/validation.csv",
+    "train": "audioCorpus/train.csv",
+}
+
+_PROMPTS_AUTOMATIC_URLS = {
+    "dev": "automatic/validation.csv",
+    "train": "automatic/train.csv",
+}
+
+_PROMPTS_TEST_URLS = {
+    "test": "test/test.csv",
+}
+
+_ARCHIVES_PROSODIC = {
+    "dev": "prosodic/audios.tar.gz",
+    "train": "prosodic/audios.tar.gz",
+}
+
+_ARCHIVES_AUDIO_CORPUS = {
+    "dev": "audioCorpus/audios.tar.gz",
+    "train": "audioCorpus/audios.tar.gz",
+}
+
+_ARCHIVES_AUTOMATIC = {
+    "train": "automatic/nurc_cm_automatic_segmented_audios.zip",
+    "dev": "automatic/nurc_cm_automatic_segmented_audios.zip",
+}
+
+_ARCHIVES_TEST = {
+    "test": "test/test.zip",
+}
+
+_PATH_TO_CLIPS = {
+    "dev": "",
+    "train": "",
+    "test": "",
+}
+
+
+class NurcSPConfig(BuilderConfig):
+    def __init__(self, prompts_type, **kwargs):
+        super().__init__(**kwargs)
+        self.prompts_type = prompts_type
+
+
+class NurcSPDataset(GeneratorBasedBuilder):
+    BUILDER_CONFIGS = [
+        NurcSPConfig(name="audioCorpus", description="Audio Corpus audio prompts", prompts_type="audioCorpus"),
+        NurcSPConfig(name="prosodic", description="Prosodic audio prompts", prompts_type="prosodic"),
+        NurcSPConfig(name="automatic", description="Automatic audio prompts", prompts_type="automatic"),
+        NurcSPConfig(name="test", description="Test audio prompts", prompts_type="test"),
+    ]
+
+    def _info(self):
+        if self.config.name == "prosodic":
+            return DatasetInfo(
+                features=datasets.Features(
+                    {
+                        "path": datasets.Value("string"),
+                        "name": datasets.Value("string"),
+                        "speaker": datasets.Value("string"),
+                        "start_time": datasets.Value("string"),
+                        "end_time": datasets.Value("string"),
+                        "normalized_text": datasets.Value("string"),
+                        "text": datasets.Value("string"),
+                        "duration": datasets.Value("string"),
+                        "type": datasets.Value("string"),
+                        "year": datasets.Value("string"),
+                        "gender": datasets.Value("string"),
+                        "age_range": datasets.Value("string"),
+                        "total_duration": datasets.Value("string"),
+                        "quality": datasets.Value("string"),
+                        "theme": datasets.Value("string"),
+                        "audio": datasets.Audio(sampling_rate=16_000, mono=True),
+                    }
+                )
+            )
+        elif self.config.name == "audioCorpus":
+            return DatasetInfo(
+                features=datasets.Features(
+                    {
+                        "audio_name": datasets.Value("string"),
+                        "file_path": datasets.Value("string"),
+                        "text": datasets.Value("string"),
+                        "start_time": datasets.Value("string"),
+                        "end_time": datasets.Value("string"),
+                        "duration": datasets.Value("string"),
+                        "quality": datasets.Value("string"),
+                        "speech_genre": datasets.Value("string"),
+                        "speech_style": datasets.Value("string"),
+                        "variety": datasets.Value("string"),
+                        "accent": datasets.Value("string"),
+                        "sex": datasets.Value("string"),
+                        "age_range": datasets.Value("string"),
+                        "num_speakers": datasets.Value("string"),
+                        "speaker_id": datasets.Value("string"),
+                        "audio": datasets.Audio(sampling_rate=16_000, mono=True),
+                    }
+                )
+            )
+        elif self.config.name == "automatic":
+            return DatasetInfo(
+                features=datasets.Features(
+                    {
+                        "path": datasets.Value("string"),
+                        "name": datasets.Value("string"),
+                        "speaker": datasets.Value("string"),
+                        "start_time": datasets.Value("string"),
+                        "end_time": datasets.Value("string"),
+                        "text": datasets.Value("string"),
+                        "duration": datasets.Value("string"),
+                        "audio": datasets.Audio(sampling_rate=16_000, mono=True),
+                    }
+                )
+            )
+        elif self.config.name == "test":
+            return DatasetInfo(
+                features=datasets.Features(
+                    {
+                        "path": datasets.Value("string"),
+                        "name": datasets.Value("string"),
+                        "speaker": datasets.Value("string"),  
+                        "start_time": datasets.Value("string"),
+                        "end_time": datasets.Value("string"),
+                        "text": datasets.Value("string"),
+                        "duration": datasets.Value("string"),
+                        "audio": datasets.Audio(sampling_rate=16_000, mono=True),                  
+                    }
+                )
+            )
+
+
+    def _split_generators(self, dl_manager):
+        if self.config.prompts_type == "prosodic":
+            prompts_urls = _PROMPTS_PROSODIC_URLS
+            archive_link = _ARCHIVES_PROSODIC
+        elif self.config.prompts_type == "audioCorpus":
+            prompts_urls = _PROMPTS_AUDIO_CORPUS_URLS
+            archive_link = _ARCHIVES_AUDIO_CORPUS
+        elif self.config.prompts_type == "automatic":
+            prompts_urls = _PROMPTS_AUTOMATIC_URLS
+            archive_link = _ARCHIVES_AUTOMATIC
+        elif self.config.prompts_type == "test":
+            prompts_urls = _PROMPTS_TEST_URLS
+            archive_link = _ARCHIVES_TEST
+        else:
+            return
+
+        prompts_path = dl_manager.download(prompts_urls)
+        archive = dl_manager.download(archive_link)
+
+        if self.config.prompts_type == "prosodic" or self.config.prompts_type == "audioCorpus":
+            return [
+                SplitGenerator(
+                    name=Split.VALIDATION,
+                    gen_kwargs={
+                        "prompts_path": prompts_path["dev"],
+                        "path_to_clips": _PATH_TO_CLIPS["dev"],
+                        "audio_files": dl_manager.iter_archive(archive["dev"]),
+                        "split_name": "validation"
+                    }
+                ),
+                SplitGenerator(
+                    name=Split.TRAIN,
+                    gen_kwargs={
+                        "prompts_path": prompts_path["train"],
+                        "path_to_clips": _PATH_TO_CLIPS["train"],
+                        "audio_files": dl_manager.iter_archive(archive["train"]),
+                        "split_name": "train"
+                    }
+                ),
+            ]
+        elif self.config.prompts_type == "automatic":
+            return [
+                SplitGenerator(
+                    name=Split.VALIDATION,
+                    gen_kwargs={
+                        "prompts_path": prompts_path["dev"],
+                        "path_to_clips": _PATH_TO_CLIPS["dev"],
+                        "audio_files": dl_manager.iter_archive(archive["dev"]),
+                        "split_name": "validation"
+                    }
+                ),
+                SplitGenerator(
+                    name=Split.TRAIN,
+                    gen_kwargs={
+                        "prompts_path": prompts_path["train"],
+                        "path_to_clips": _PATH_TO_CLIPS["train"],
+                        "audio_files": dl_manager.iter_archive(archive["train"]),
+                        "split_name": "train"
+                    }
+                ),
+            ]
+        elif self.config.prompts_type == "test":
+            return[
+                SplitGenerator(
+                    name=Split.TEST,
+                    gen_kwargs={
+                        "prompts_path": prompts_path["test"],
+                        "path_to_clips": _PATH_TO_CLIPS["test"],
+                        "audio_files": dl_manager.iter_archive(archive["test"]),
+                        "split_name": "test"
+                    }
+                ),
+            ]
+
+    def _generate_examples(self, prompts_path, path_to_clips, audio_files, split_name):
+        examples = {}
+        csv_paths = []
+
+        with open(prompts_path, "r", encoding="utf-8") as f:
+            if self.config.prompts_type == "test":
+                csv_reader = csv.DictReader(f, delimiter=";")  # Explicitly set delimiter
+            else:
+                csv_reader = csv.DictReader(f)
+                
+            if self.config.prompts_type == "prosodic":
+                for row in csv_reader:
+                    file_path = Path(row['path']).as_posix()
+                    examples[file_path] = {
+                        "path": row['path'],
+                        "name": row['name'],
+                        "speaker": row['speaker'],
+                        "start_time": row['start_time'],
+                        "end_time": row['end_time'],
+                        "normalized_text": row['normalized_text'],
+                        "text": row['text'],
+                        "duration": row['duration'],
+                        "type": row['type'],
+                        "year": row['year'],
+                        "gender": row['gender'],
+                        "age_range": row['age_range'],
+                        "total_duration": row['total_duration'],
+                        "quality": row['quality'],
+                        "theme": row['theme'],
+                    }
+                    csv_paths.append(file_path)
+            elif self.config.prompts_type == "audioCorpus":
+                for row in csv_reader:
+                    file_path = Path(row['file_path']).as_posix()
+                    examples[file_path] = {
+                        "audio_name": row['audio_name'],
+                        "file_path": row['file_path'],
+                        "text": row['text'],
+                        "start_time": row['start_time'],
+                        "end_time": row['end_time'],
+                        "duration": row['duration'],
+                        "quality": row['quality'],
+                        "speech_genre": row['speech_genre'],
+                        "speech_style": row['speech_style'],
+                        "variety": row['variety'],
+                        "accent": row['accent'],
+                        "sex": row['sex'],
+                        "age_range": row['age_range'],
+                        "num_speakers": row['num_speakers'],
+                        "speaker_id": row['speaker_id'],
+                    }
+                    csv_paths.append(file_path)
+            elif self.config.prompts_type == "automatic":
+                for row in csv_reader:
+                    file_path = Path(row['path']).as_posix()
+                    examples[file_path] = {
+                        "path": row['path'],
+                        "name": row['name'],
+                        "speaker": row['speaker'],
+                        "start_time": row['start_time'],
+                        "end_time": row['end_time'],
+                        "text": row['text'],
+                        "duration": row['duration'],
+                    }
+                    csv_paths.append(file_path)
+            elif self.config.prompts_type == "test":
+                for row in csv_reader:
+                    file_path = Path(row['path']).as_posix()
+                    examples[file_path] = {
+                        "path": row['path'],
+                        "name": row['name'],
+                        "speaker": row['speaker'],
+                        "start_time": row['start_time'],
+                        "end_time": row['end_time'],
+                        "text": row['text'],
+                        "duration": row['duration'],
+                    }
+                    csv_paths.append(file_path)
+
+        id_ = 0
+        for path, f in audio_files:
+            path = Path(path).as_posix()
+            if path.startswith(path_to_clips) and path in examples:
+                audio = {"path": path, "bytes": f.read()}
+                yield id_, {**examples[path], "audio": audio}
+                id_ += 1
+
+
+if __name__ == "__main__":
+    from datasets import DatasetDict, DatasetInfo
+    import json
+    from pathlib import Path
+
+    infos = {}
+
+    for cfg in ["prosodic", "audioCorpus", "automatic", "test"]:
+        # Load parquet files into a DatasetDict
+        parquet_files = list(Path(f"parquet/{cfg}").glob("*.parquet"))
+        ds = DatasetDict({
+            f.stem.split("_")[-1]:  # e.g., "train", "validation", "test"
+            DatasetDict.from_parquet(str(f)) for f in parquet_files
+        })
+
+        # Build info object
+        info = DatasetInfo.from_dataset_dict(ds)
+
+        # Store in dict
+        infos[cfg] = info.to_dict()
+
+    # Save to dataset_infos.json
+    with open("dataset_infos.json", "w", encoding="utf-8") as f:
+        json.dump(infos, f, indent=2, ensure_ascii=False)
+
+    '''
+    for cfg in ["prosodic", "audioCorpus", "automatic", "test"]:
+        builder = NurcSPDataset(config_name=cfg)
+        builder.download_and_prepare()
+        ds: DatasetDict = builder.as_dataset()
+
+        out_dir = Path(f"parquet/{cfg}")
+        out_dir.mkdir(parents=True, exist_ok=True)
+
+        for split, dataset in ds.items():
+            dataset.to_parquet(out_dir / f"{cfg}_{split}.parquet")
     '''