Update dataset loading script

jeli-asr.py

@@ -14,8 +14,8 @@ See the License for the specific language governing permissions and
 limitations under the License.
 """
 # jeli_asr.py
-import os
 import csv
+import os
 
 import datasets
 from datasets import Split, SplitGenerator
@@ -37,7 +37,7 @@ the cleaned subset of Oza's Bambara-ASR with the Jeli-ASR-RMAI subset.
 
 _HOMEPAGE = "https://huggingface.co/datasets/RobotsMali/jeli-asr"
 _LICENSE = "CC-BY-4.0"
-
+VERSION = datasets.Version("1.0.0")
 
 class JeliASRConfig(datasets.BuilderConfig):
     """BuilderConfig for different subsets of the Jeli-ASR dataset."""
@@ -50,17 +50,17 @@ class JeliASR(datasets.GeneratorBasedBuilder):
     BUILDER_CONFIGS = [
         JeliASRConfig(
             name="jeli-asr-rmai",
-            version=datasets.Version("1.0.0"),
+            version=VERSION,
             description="Load only the Jeli-ASR RMAI subset (files under jeli-asr-rmai/)",
         ),
         JeliASRConfig(
             name="bam-asr-oza",
-            version=datasets.Version("1.0.0"),
+            version=VERSION,
             description="Load only the Bam-ASR Oza subset (files under bam-asr-oza/)",
         ),
         JeliASRConfig(
             name="jeli-asr",
-            version=datasets.Version("1.0.0"),
+            version=VERSION,
             description="Combined data: jeli-asr-rmai + bam-asr-oza (all rows).",
         ),
     ]
@@ -73,7 +73,7 @@ class JeliASR(datasets.GeneratorBasedBuilder):
             description=_DESCRIPTION,
             features=datasets.Features(
                 {
-                    # We'll load the audio from local paths (relative to data_dir + file_name).
+                    # We'll load the audio from local paths.
                     "audio": datasets.Audio(sampling_rate=16_000),
                     "duration": datasets.Value("float32"),
                     "bam": datasets.Value("string"),
@@ -88,15 +88,15 @@ class JeliASR(datasets.GeneratorBasedBuilder):
 
     def _split_generators(self, dl_manager):
         """
-        We assume metadata.csv is in data_dir. We'll generate two splits:
-        - TRAIN: file paths containing "/train/"
-        - TEST: file paths containing "/test/"
+        We'll generate two splits by scanning metadata.csv for rows
+        containing '/train/' or '/test/' in the file path.
         """
-        metadata_path = "metadata.csv"
+        # If the user doesn't pass data_dir, self.config.data_dir might be empty or None,
+        # so let's handle that gracefully:
+        base_dir = self.config.data_dir or ""
 
-        # We do not need to call dl_manager.download_and_extract here because
-        # the audio + CSV are already on the Hub. The Datasets library will handle
-        # fetching them automatically if not present locally.
+        # metadata.csv is assumed to be in the repo root (or base_dir if there's a subfolder).
+        metadata_path = os.path.join(base_dir, "metadata.csv")
 
         return [
             SplitGenerator(
@@ -104,6 +104,7 @@ class JeliASR(datasets.GeneratorBasedBuilder):
                 gen_kwargs={
                     "metadata_path": metadata_path,
                     "split": "train",
+                    "base_dir": base_dir
                 },
             ),
             SplitGenerator(
@@ -111,50 +112,47 @@ class JeliASR(datasets.GeneratorBasedBuilder):
                 gen_kwargs={
                     "metadata_path": metadata_path,
                     "split": "test",
+                    "base_dir": base_dir
                 },
             ),
         ]
 
-    def _generate_examples(self, metadata_path, split):
+    def _generate_examples(self, metadata_path, split, base_dir):
         """
-        Read metadata.csv row by row. 
-        Filter rows by:
+        Read metadata.csv row-by-row. Filter by:
           1) Config name (jeli-asr-rmai, bam-asr-oza, or jeli-asr).
-          2) Split substring (train/test).
-        Yield each example with features: audio, duration, bam, french.
+          2) 'train' vs 'test' in the file path.
+        Yield each example with audio, duration, bam, and french.
         """
         with open(metadata_path, "r", encoding="utf-8") as f:
             reader = csv.DictReader(f)
             idx = 0
             for row in reader:
-                file_path = row["file_name"]  # e.g. "bam-asr-oza/train/split1/oza75-bam-asr-0.wav"
+                file_path = row["file_name"]  # e.g. 'bam-asr-oza/train/split1/oza75-bam-asr-0.wav'
 
-                # 1) Filter by config name
+                # Filter by config
                 if self.config.name == "jeli-asr-rmai":
-                    # Keep only rows with 'jeli-asr-rmai/' in the file path
                     if "jeli-asr-rmai/" not in file_path:
                         continue
                 elif self.config.name == "bam-asr-oza":
-                    # Keep only rows with 'bam-asr-oza/' in the file path
                     if "bam-asr-oza/" not in file_path:
                         continue
-                # If config.name == "jeli-asr", we include all rows.
+                # If config is "jeli-asr", include everything.
 
-                # 2) Filter by split (train/test)
+                # Filter by train/test
                 if split == "train" and "/train/" not in file_path:
                     continue
                 if split == "test" and "/test/" not in file_path:
                     continue
 
-                # Construct the full path to the audio file, relative to self.config.data_dir
-                # e.g. "jeli-asr/bam-asr-oza/train/..."
-                #audio_path = os.path.join(self.config.data_dir, file_path)
+                # Build the absolute local path to the audio file (within the cloned repo)
+                audio_path = os.path.join(base_dir, file_path)
 
-                # Yield a dictionary matching the schema in _info()
                 yield idx, {
-                    "audio": file_path,
+                    "audio": audio_path,
                     "duration": float(row["duration"]),
                     "bam": row["bam"],
                     "french": row["french"],
                 }
                 idx += 1
+