dsiter 0.1.1

A library for iterating over aggregated datasets from HuggingFace and local CSV files

dsiter

ds-iter (ds is short for "dataset" and iter for "iteration") is an opinionated, declarative dataset preparation/iteration tool that helps you create an single iterator over many datasets, useful for ML/data engineering tasks.

Problem

During the training of large-scale machine learning models (especially language models or NLP tasks), we're often working with many different datasets from various sources, with very different shapes and dimensions, sometimes reaching gigabytes of data.

Failure in implementing efficient memory usage within training or data preparation scripts often leads to operating system errors, such as OOM killed.

ML/data engineers tend to write custom scripts, often with bespoke pre-processing/post-processing mutations, and sometimes entirely different implementations of the data preparation step across projects. This can lead to reproducibility issues, as different team members using slightly different scripts produce inconsistent results. Additionally, debugging becomes difficult because OOM errors or preprocessing bugs are often non-deterministic and hard to trace.

Usage

step one: create yaml configuration file

create a yaml file (by default the library looks for datasets.yaml) and start adding your Hugginface 🤗 datasets using their repoId like below:

Example: See examples/datasets.yml for a complete configuration file with various dataset types.

datasets:
  - path: facebook/recycling_the_web

you can add as many Hugginface 🤗 datasets as you want:

datasets:
  - path: facebook/recycling_the_web
  - path: Lk123/InfoSeek
  - path: ...
  - path: ...

you can also load your own custom dataset files, supported formats are:

• csv or tsv
• parquet

just pass the relative/absolute path of the dataset with path key.

datasets:
  - path: ./files/dataset_dump.csv
  - path: ./files/other_dataset.tsv
  - path: ./files/other_other_dataset_0001.parquet

for each dataset you can target an array of specific columns to be returned within iterator:

datasets:
  - path: MathLLMs/MathVision
    columns:
      - questions

  - path: community-datasets/farsi_news
    columns:
      - title
      - summary

step two: python scripting

install the library using pip

pip install dsiter

and then:

Example: See examples/example.py for a complete working example.

from dsiter import DSIterCollection

collection = DSIterCollection()

for row in collection.iter_rows():
    print(row)

Calling iter_rows() lazily streams the dataset's rows through a generator, enabling efficient iteration and processing without loading the entire dataset into memory.