Datasets:

simecek
/

Human_DNA_v0

license: other
tags:
  - biology
  - genomics
  - DNA
  - huggingscience
  - science
pretty_name: Human DNA v0
size_categories:
  - 100K<n<1M
task_categories:
  - text-generation
  - fill-mask

Dataset Card for Human DNA v0

Dataset Description

Repository: simecek/Human_DNA_v0

Dataset Summary

The Human_DNA_v0 dataset provides a large corpus of nucleotide sequences from the human genome. It is designed to serve as a foundational training resource for developing and evaluating language models in the field of genomics. The data consists of long strings representing the fundamental DNA bases: A (adenine), C (cytosine), G (guanine), and T (thymine).

This dataset is crucial for pre-training models that can learn the underlying patterns, syntax, and long-range dependencies within the human genetic code. Such "Genomic Foundation Models" have applications in predicting gene locations, understanding regulatory regions, and identifying variants associated with disease. 🧬

Supported Tasks and Leaderboards

This dataset is primarily intended for self-supervised learning tasks, similar to how models like BERT or GPT are trained on natural language.

Masked Language Modeling (Fill-Mask): Models can be trained to predict a masked or missing nucleotide in a sequence. This helps the model learn contextual information within the DNA strand.
Next-Token Prediction (Text Generation): Models can be trained to predict the next nucleotide in a sequence. This is the basis for generative models that can produce realistic DNA sequences.
Downstream Fine-tuning: Models pre-trained on this dataset can be fine-tuned for a variety of specific genomic tasks, such as promoter identification, splice site prediction, or classification of non-coding regions.

Data Fields

The dataset is provided in Parquet format and contains a single, straightforward column:

sequence: (string) A long string of characters representing a segment of the human DNA sequence. Each character is one of A, C, G, or T.

Data Splits

train: Contains 292,955 rows of DNA sequences for model training.

There are no pre-defined validation or test splits, so users should create their own splits from the training data as needed for their experiments.

Curation and Rationale

While the original curation details are sparse, this dataset was created to provide a simple, large-scale resource for applying modern NLP techniques to genomics. By formatting genomic data in a way that is immediately accessible to machine learning libraries, it lowers the barrier for ML practitioners to contribute to computational biology. It has been successfully used as a pre-training corpus for models like vesteinn/gpt2-dna.

Considerations for Use

Character-level Processing: DNA sequences are inherently character-level. Tokenization strategies are a key consideration. Simple character-level tokenizers are common, but more advanced methods like BPE (Byte-Pair Encoding) or k-mer based approaches may also be effective. A **k-mer