GenomeQA

GenomeQA tests whether a general-purpose language model can answer questions about raw nucleotide sequences, with no specialised DNA encoder and no task-specific training. Each question is English text containing one or more DNA sequences, answerable from signals such as sequence composition, motifs, and motif context.

5,200 examples — 2,600 binary-choice and 2,600 four-option — with sequences from 6 to 1,000 bp, across six task families:

1. enhancer and promoter identification;
2. splice-site identification;
3. taxonomic classification;
4. histone-mark and chromatin-state prediction;
5. transcription-factor binding-site prediction; and
6. transcription-factor motif prediction.

Source and construction

This repository repackages the benchmark introduced in GenomeQA: Benchmarking General Large Language Models for Genome Sequence Understanding (arXiv:2604.05774) by Weicai Long, Yusen Hou, Junning Feng, Houcheng Su, Shuo Yang, Donglin Xie, and Yanlin Zhang, as Parquet with the extra subsets listed below. We use it to evaluate Carbon-X.

The paper constructs GenomeQA from established genomic resources:

Task family	Primary source	Sequence length
Enhancer and promoter identification	EPDnew and ENCODE SCREEN	400 bp
Splice-site identification	Nucleotide Transformer downstream tasks	600 bp
Taxonomic classification	NCBI RefSeq	1,000 bp
Histone-mark prediction	Nucleotide Transformer downstream tasks	1,000 bp
TFBS prediction	ENCODE and JASPAR	100 bp
TF motif prediction	ENCODE and JASPAR	6–20 bp

The authors standardize sequence windows, remove overlapping examples to reduce ambiguity, exclude sequences with ambiguous nucleotide bases, and convert the resulting annotations into binary-choice and multiple-choice question templates. Refer to the paper for the full construction procedure and source-database citations.

Configurations

All configurations contain a single train split. The full benchmark configurations are intentionally not equal in size across subtasks: each contains 500 examples for the first five task families and 100 motif examples. Benchmark results should therefore report per-subtask accuracy and use an unweighted mean across the six subtasks when a single headline score is needed.

Configuration	Rows	Meaning
binary	2,600	Full benchmark in binary-choice format, with two ordered options per question.
mcq	2,600	Full benchmark in four-option multiple-choice format.
mcq_carbon	2,600	MCQ benchmark with the 500 broad taxonomic-classification (3_species_cls) rows replaced by 500 stratified within-eukaryote eukaryote_cls rows.
mcq_carbon_600	600	Stratified 100-row sample from each mcq_carbon subtask.
binary_600	600	Lower-cost binary subset with 100 examples from each of the six subtasks.
mcq_600	600	Lower-cost MCQ subset with 100 examples from each of the six subtasks.
binary_eukaryote	170	Binary 3_species_cls questions whose target organism is eukaryotic.
binary_prokaryote	150	Binary 3_species_cls questions whose target organism is prokaryotic.
mcq_eukaryote	177	MCQ 3_species_cls questions whose target organism is eukaryotic.
mcq_prokaryote	165	MCQ 3_species_cls questions whose target organism is prokaryotic.
eukaryote_cls	1,000	Eukaryote classification within eukaryotes: all four options are eukaryotic taxa, so the model must tell eukaryotic lineages apart rather than separate eukaryote from prokaryote. Auxiliary set, stored separately from the original 5,200-example benchmark.

The mcq_carbon configuration preserves the other mcq rows and inserts 500 eukaryote_cls questions at the original species-classification position. From source revision ae30ecec16a7669f342f7d87439552142cdebe40, seed 42 selects 300 basic_class, 100 reverse_search, and 100 odd_one_out rows, sampling approximately half of each taxon label within each question type. See scripts/create_mcq_carbon.py.

The mcq_carbon_600 configuration samples without replacement from mcq_carbon at revision c0a7d5b8ccfe4c0d64bc3ff5dd171ff4365c16f6 using seed 42. It retains 100 rows per subtask and preserves the mix of question dimensions within each subtask. Within dimensions, it samples proportionally by target label where available; eukaryote odd-one-out questions are also stratified by majority taxon, and answer position is the final stratum. The 100 motif rows are all retained. See scripts/create_mcq_carbon_600.py.

The eukaryote_cls configuration uses the mcq feature schema, including JSON-serialized metadata.

The binary_600 and mcq_600 configurations were sampled without replacement using seed 42 from source revision a71ab7e0e72b8eca7acd4475744dccd0159b20be. The 100 selected examples from each subtask were then combined and shuffled so the rows are not stored in subtask blocks. These subsets are balanced by task family, but not necessarily by every internal label or question dimension. See scripts/create_600_subsets.py for the reproducible generation script.

The organism-specific configurations contain only rows from the 3_species_cls subtask and preserve their source order. The target organism is read from metadata: true for binary 1_identity and MCQ basic_class questions, target for MCQ reverse_search questions, and minority for MCQ odd_one_out questions because the minority class is the sequence the model must identify. Eukaryote and prokaryote subsets are disjoint within each answer format; virus rows are not included. See scripts/create_organism_subsets.py for the reproducible filtering script.

Data fields

Field	Description
id	Integer identifier within the source subtask.
subtask	Task-family identifier used for grouped evaluation.
question	Natural-language question containing the relevant DNA sequence or sequences.
options	Ordered list of candidate answers.
answer_index	Zero-based index of the correct entry in options.
answer_text	Correct answer text. For sequence-selection questions, this may itself be a DNA sequence.
metadata	Serialized task-specific metadata describing the question dimension and source label.

Usage

from datasets import load_dataset


binary = load_dataset("HuggingFaceBio/genomeqa", "binary", split="train")
representative_mcq = load_dataset("HuggingFaceBio/genomeqa", "mcq_600", split="train")
eukaryote_mcq = load_dataset("HuggingFaceBio/genomeqa", "mcq_eukaryote", split="train")

Citation

@article{long2026genomeqa,
  title={GenomeQA: Benchmarking General Large Language Models for Genome Sequence Understanding},
  author={Long, Weicai and Hou, Yusen and Feng, Junning and Su, Houcheng and Yang, Shuo and Xie, Donglin and Zhang, Yanlin},
  journal={arXiv preprint arXiv:2604.05774},
  year={2026}
}

Reproducing

To rebuild and publish the normalized eukaryote_cls configuration from the pinned source revision, run:

python scripts/create_eukaryote_cls.py --source-revision 18f68fabbcd43e8c227869a833027c969b39d7fe --output-dataset HuggingFaceBio/genomeqa --publish

The script preserves the source rows, converts metadata from Python dictionary strings to JSON, and casts the columns to the mcq feature schema. It publishes the Parquet file, this card, and the builder script in one commit.

mcq_carbon

To rebuild and publish mcq_carbon from the pinned source revision, run:

python scripts/create_mcq_carbon.py --source-revision ae30ecec16a7669f342f7d87439552142cdebe40 --output-dataset HuggingFaceBio/genomeqa --publish

The script publishes the Parquet file, this card, and the builder script together.

mcq_carbon_600

To rebuild and publish mcq_carbon_600 from the pinned source revision, run:

python scripts/create_mcq_carbon_600.py --source-revision c0a7d5b8ccfe4c0d64bc3ff5dd171ff4365c16f6 --output-dataset HuggingFaceBio/genomeqa --seed 42 --publish

The script publishes the Parquet file, this card, and the builder script together.