---
license: mit
task_categories:
  - text-generation
tags:
  - long-sequence-generation
  - training-data
  - llm-training
dataset_info:
  features:
  - name: input_ids
    sequence: int32
  splits:
  - name: train
    num_bytes: 1126604624
    num_examples: 27553
  download_size: 521687503
  dataset_size: 1126604624
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---

This dataset provides training data for accelerating ultra-long sequence generation with large language models (LLMs).  It's 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).

The data is derived from the PG-19 dataset, filtering out sequences smaller than 8K tokens to focus on ultra-long sequences relevant for the TokenSwift method described in the paper.

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