Datasets:

iliasslasri
/

tokenized-OLMoE-mix

license: apache-2.0
task_categories:
  - text-generation
language:
  - en
tags:
  - moE
  - olmoe
  - pretraining
  - allenai
pretty_name: Tokenized OLMoE Mix
size_categories:
  - 1B<n<10B

Dataset Card for Tokenized OLMoE Mix

Dataset Summary

This dataset contains pre-tokenized training and evaluation data designed for training custom small-scale OLMoE (Mixture-of-Experts) models.

The data is sourced primarily from the official AI2 Dolma 1 and C4 datasets and was curated to run ablation studies and reproduction experiments related to the OLMoE Technical Paper (arXiv:2409.02060). It is provided in .npy format, having been pre-tokenized using the allenai/gpt-neox-olmo-dolma-v1_5 tokenizer.

Dataset Structure

The dataset currently totals 7.38 GB and is split into three parts for easier handling:

part-0-00000.npy (2.51 GB)
part-0-00001.npy (4.29 GB)
part-0-00002.npy (574 MB)

Data Sources & Composition

Our training mix consists of approximately 4.7 Billion tokens in total, built from the following sources:

1. Training Data (3.689B tokens)

Source: A Wikipedia subset from Dolma 1.
Original HF Dataset: allenai/OLMoE-mix-0924
Command used to fetch raw data:

wget -O data/wiki-001.json.gz "[https://huggingface.co/datasets/allenai/OLMoE-mix-0924/resolve/main/data/wiki/wiki-0001.json.gz?download=true](https://huggingface.co/datasets/allenai/OLMoE-mix-0924/resolve/main/data/wiki/wiki-0001.json.gz?download=true)"