pythia-training-evals.py

_DESCRIPTION="""\
    Dataset for storing training evaluations of pythia models, e.g. loss, perplexity
"""

import datasets
import json

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. 
        """
        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)])

        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