metadata

license: mit
task_categories:
  - text-generation
tags:
  - biology
  - genomics
  - long-context
configs:
  - config_name: bacteria
    data_files:
      - split: test
        path: bacteria/test.parquet
  - config_name: eukaryote
    data_files:
      - split: test
        path: eukaryote/test.parquet
  - config_name: others
    data_files:
      - split: test
        path: others/test.parquet

Next K-mer Prediction

Abouts

The Next K-mer Prediction task is a zero-shot evaluation method introduced in the GENERator paper to assess the quality of pretrained models. It involves inputting a sequence segment into the model and having it predict the next K base pairs. The predicted sequence is then compared to the actual sequence to assess accuracy.

Sequence: The input sequence has a maximum length of 96k base pairs (bp). You can control the number of input tokens by applying left truncation.
Label: The next 128 bp immediately following the end of the input sequence.

Note: Prediction time may increase significantly for longer input sequences. It is strongly recommended to begin testing with a smaller number of input tokens to optimize performance.

How to use

from datasets import load_dataset

datasets = load_dataset("GenerTeam/next-kmer-prediction", "eukaryote") # or "bacteria" or "others"

Citation

@misc{wu2025generator,
    title={GENERator: A Long-Context Generative Genomic Foundation Model},
    author={Wei Wu and Qiuyi Li and Mingyang Li and Kun Fu and Fuli Feng and Jieping Ye and Hui Xiong and Zheng Wang},
    year={2025},
    eprint={2502.07272},
    archivePrefix={arXiv},
    primaryClass={cs.CL},
    url={https://arxiv.org/abs/2502.07272},
}