README.md

metadata

license: apache-2.0
task_categories:
  - text-generation
tags:
  - biology
  - genomics
  - dna
size_categories:
  - 10M<n<100M

carbon-cpu-enriched-sequences

A CPU-enriched subset of the carbon pretraining corpus (eukaryote_generator), combining original source fields with normalized sequences and row-level features for quality analysis, GPU enrichment and embedding generation.

---

Information of Features

Feature	Type	Description
record_id	string	NCBI Identifier linking the row back to the source genomic record. It provides the primary record-level identity.
begin_of_sequence	string	Beginning-of-sequence special token marking the start of a sequence record, e.g. <s>.
end_of_sequence	string	End-of-sequence special token marking the end of a sequence record, e.g. </s>.
begin_of_gene	string	Marker identifying the beginning of a gene/feature segment, e.g. <bog>.
end_of_gene	string	Marker identifying the end of a gene/feature segment, e.g. <eog>.
gene_type	string	Annotation describing the type of genomic feature. In the observed data, <cds> represents a coding sequence and <fng> represents another gene-feature category.
species_type	string	Broad biological category of the source sequence. The current data is represented as Eukaryota.
strand	string	DNA strand/orientation associated with the sequence, represented by + or -.
sequence	string	The actual nucleotide sequence represented by the record. This is the primary biological sequence field.
molecule_type	string	Molecular type of the sequence; the current data is represented as DNA.
topology	string	Structural topology of the molecule, such as linear.
taxonomy	string	Hierarchical taxonomic lineage of the sequence, with ranks represented as a semicolon-separated lineage, e.g. Eukaryota;Fungi;...;Agaricus.
start	int64	Starting position of the record within the pre-trained Carbon corpus. It is a corpus coordinate, not necessarily a genomic coordinate.
end	int64	Ending position of the record within the processed/tokenized Carbon corpus.
gc_content	float64	Fraction of nucleotides that are G or C: (G + C) / sequence_length. It measures GC composition.
gc_skew	float64	Measures asymmetry between G and C: (G - C) / (G + C). Positive values indicate relative G enrichment; negative values indicate relative C enrichment.
sequence_length	int64	Number of nucleotide bases in the sequence.
gene_length	int64	Length of the associated gene/feature sequence. For the current records it commonly corresponds to sequence_length.
shannon_entropy	float64	Sequence-complexity measure based on nucleotide-frequency distribution.
kmer_frequency_vector	list	A 64-element vector containing normalized frequencies of all possible DNA 3-mers (4³ = 64). It captures local sequence-composition patterns.
strand_normalized_sequence	string	Sequence transformed into a consistent orientation so that strand direction does not create artificial differences in downstream analysis.
taxonomy_domain	string	Broadest taxonomic/domain-level classification extracted from the taxonomy hierarchy.
taxonomy_depth	int64	Number of taxonomic levels represented in the lineage. It provides a compact measure of how deeply the sequence is classified.
is_coding_region	bool	Boolean indicator of whether the record is classified as a coding sequence. The pipeline identifies <cds> as the coding gene type.
qc_flag	string	Quality-control status assigned to the record, indicating whether it passes the pipeline's sequence-level QC criteria.

The dataset contains both the original Carbon fields and derived enrichment features.

---

Dataset Summary

carbon-cpu-enriched-sequences provides a reusable CPU-side canonical representation of the Carbon genomic corpus before computationally expensive GPU enrichment.

The original Carbon source corpus contains approximately 46.3 million records. This release contains 32,410,000 records, corresponding to approximately 70% of the source row population. The release was selected as a practical balance between corpus representation, CPU processing cost, storage requirements, and downstream GPU compute availability.

The dataset preserves the original Carbon source representation while adding CPU-derived enrichment features:

Original Carbon columns
        +
CPU-derived enrichment columns
        ↓
CPU-enriched Parquet dataset

The current materialized release contains approximately 25 columns, consisting of approximately 14 source/raw columns and 11 derived enrichment columns. The exact schema should be inspected from the Parquet files, as future enrichment revisions may add or modify derived fields.

CPU enrichment includes sequence-level and metadata-derived information such as sequence characteristics, coding status, strand/orientation information, and taxonomy depth. Coding sequences are identified using the observed Carbon representation <cds>.

The dataset is stored as Apache Parquet with Zstandard compression, using approximately 50,000 rows per shard.

Intended Uses

This dataset is intended for research and data-engineering workflows involving genomic sequence corpora, including:

• Genomic data quality and validation analysis
• Sequence-distribution and statistical analysis
• CPU-side preprocessing research
• Dataset sampling and cohort construction
• GPU enrichment experiments
• Embedding generation
• Nearest-neighbor and clustering analysis
• Genomic foundation-model data engineering
• Benchmarking large-scale data-processing pipelines
• Construction of downstream enriched datasets

For large-scale datasets, streaming access is recommended rather than loading the entire dataset into memory.

---