---
dataset_info:
  features:
  - name: input_ids
    sequence: int32
  splits:
  - name: train
    num_bytes: 1083419904
    num_examples: 27352
  download_size: 487720508
  dataset_size: 1083419904
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
task_categories:
- text-generation
---

This repository contains the PG-19 training dataset, used in the paper [From Hours to Minutes: Lossless Acceleration of Ultra Long Sequence Generation up to 100K Tokens](https://hf.co/papers/2502.18890).
Data larger than 8K tokens are filtered out according to different tokenizers.

Code: https://github.com/bigai-nlco/TokenSwift