Setting up data loading script

Data loading script first attempt that simply returns the evaluation json for each checkpoint. Data splits correspond to the different model sizes.

pythia-training-evals.py

@@ -0,0 +1,74 @@
+"""
+    Dataset for storing training evaluations of pythia models, e.g. loss, perplexity
+"""
+
+class PythiaTrainingEvals(datasets.GeneratorBasedBuilder):
+    
+    MODEL_SIZES = [ 
+        "70m", 
+        "160m", 
+        "410m",
+        "1.4b",
+        "2.8b",
+    ]
+
+    BUILDER_CONFIGS = []
+    for model_size in MODEL_SIZES:
+        BUILDER_CONFIGS.extend([
+            datasets.BuilderConfig(
+                name=f"{model_size}",
+                description=f"Dataset of pythia training evaluation metrics for pythia model size: {model_size}",
+                version="1.0.0",
+            ),
+        ])
+
+    def _info(self):
+        return datasets.DatasetInfo(
+            description=_DESCRIPTION,
+        )
+
+
+    def _split_generators(self, dl_manager: datasets.DownloadManager):
+        """ 
+        Returns data for different splits - we define a split as a model size. 
+        """
+
+        to_download_files = []
+        model_size = self.config.name.split("__")[0]
+
+        for checkpoint_step in checkpoint_steps:
+            to_download_files.append(f"./models/{model_size}/checkpoint_{checkpoint_step}/evals.json")
+
+        downloaded_files = dl_manager.download_and_extract(to_download_files)
+
+        return [
+            datasets.SplitGenerator(
+                name='default',
+                gen_kwargs={
+                    "filepaths": downloaded_files,
+                }
+            )  
+        ]
+
+    def _generate_examples(self, filepaths):
+        """
+        Yields examples from each file in filepaths that are stored as jsons
+        with the evaluation metrics for a given checkpoint step. 
+        """
+
+        checkpoint_steps = [0, 1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 1000, ]
+        checkpoint_steps.extend([3000 + (i * 10000) for i in range(0, 15)])
+
+        # the filepaths should be a list of filepaths 
+        if isinstance(filepaths, str):
+            filepaths = [filepaths]
+
+        for idx, filepath in enumerate(filepaths):
+            with open(filepath, 'rb') as f:
+                data = json.load(f)
+
+            record = { 
+                "checkpoint_step": checkpoint_steps[idx],
+                **data
+            } 
+           yield idx,  record