Dataset Viewer
Auto-converted to Parquet Duplicate
id
stringlengths
67
119
aspect
stringclasses
7 values
question_type
stringclasses
2 values
question
stringclasses
172 values
dna_sequence
stringlengths
262
124k
choices
listlengths
10
10
prompt_template
stringlengths
285
2.53k
answer_indices
listlengths
1
4
answer_letters
listlengths
1
4
answer_texts
listlengths
1
4
dna_provenance
dict
annotation_provenance
dict
distractor_provenance
dict
release_provenance
dict
go_provenance
dict
geometry_provenance
stringlengths
236
2.99k
GCF_000001405.40::gene-ZNF296::NC_000019.10:45::synthesis.aspect.biological_process
biological_process
single_choice
Which one of the following Gene Ontology biological-process terms is a supported annotation for the gene contained within this stretch of DNA?
GCCCTCGCCTCAGCCCCTGCCCTGGCCTCGCCCGCACTAGGAACTGGGAGCGCAGCTTTCTGTGCCAGTCCTGCTGCGCCCCAGGCACTGGGGGCTGGCCGTTCCCTCACCCCCTCACTCCAAACCGGTCTGTCCAGCCGTCCACCCCACGGAGGGGAAGCCTCCTGCCCCACCTCATGCACCCCAAGAAAGCCAGAGTCCAGACGGGTGTAAATAAAAGGGAAAATTTACTTGGAGAAGGCGGGCAAAAACGAGGAGGTCAATGGGTGTTGGCAGCGGTACCAGGGACAGTGAGGGGGGCTTTCCTGGGCTCAGGCCTC...
[ "regulation of translation in response to endoplasmic reticulum stress", "copper ion transport", "dTMP salvage", "negative regulation of adenylate cyclase-activating adrenergic receptor signaling pathway", "positive regulation of natural killer cell chemotaxis", "mitochondrial tRNA wobble position uridine...
Which one of the following Gene Ontology biological-process terms is a supported annotation for the gene contained within this stretch of DNA? <dna_placeholder> Choose the best answer: A) regulation of translation in response to endoplasmic reticulum stress B) copper ion transport C) dTMP salvage D) negative regulation...
[ 7 ]
[ "H" ]
[ "regulation of transcription by RNA polymerase II" ]
{ "organism": "Homo sapiens", "taxonomic_group": "vertebrate_mammalian", "assembly_accession": "GCF_000001405.40", "sequence_accession": "NC_000019.10", "gene_id": "gene-ZNF296", "strand": -1, "chunk_index": 45, "gene_start_chunk": 71499, "gene_end_chunk": 76478, "dna_start_chunk": 71289, "dna_end...
{ "canonical_answer": "regulation of transcription by RNA polymerase II", "entities": [ "regulation of transcription by RNA polymerase II" ], "go_ids": [ "GO:0006357" ], "localization_labels": [], "sources": [ "goa_uniprot_all_gaf_20260410" ], "quality": 1, "source_task_id": "synthesis.a...
{ "method": "atomic_go_v5", "source_gene_ids": [ "gene-EIF2S1", "gene-Slc30a1", "gene-TK1", "gene-Pde4b", "gene-CCL7", "gene-TRMU", "gene-Synm", "gene-EXOSC9", "gene-ZRANB1" ], "source_sample_ids": [ "1c0132e9e665c3164589d255e4bc39bd99889a286f270707453faa511a852984", ...
{ "dataset_name": "GeneAnnotationBench", "source_sample_id": "GCF_000001405.40::gene-ZNF296::NC_000019.10:45::synthesis.aspect.biological_process", "source_task_id": "synthesis.aspect.biological_process", "benchmark_task_id": "bench.aspect_biological_process_mc", "source_split": "eval", "row_seed": "1557151...
{ "schema_version": 1, "policy_version": "atomic-go-v5.0", "selection_seed": 20260804, "gold_go_id": "GO:0006357", "gold_go_name": "regulation of transcription by RNA polymerase II", "gold_go_depth": 9, "gold_prevalence": 0.012554147376149783, "gold_information_content": 4.377704199777833, "gold_evide...
{"biological_review": [], "dna_changed": false, "holds": [], "policy": "gab-conditional-targeting-v1", "source_record_sha256": "53b4b4ab0bdb9273ecf9b4ddd0fa847a95940438118226da0969170b4120c89f", "targeting": {"mode": "none", "text": ""}}
GCF_000001405.40::gene-OPA3::NC_000019.10:45::synthesis.aspect.biological_process
biological_process
single_choice
Which one of the following Gene Ontology biological-process terms is a supported annotation for the gene contained within this DNA span?
AGTTTCTGATTTGGGAAAGGGCTCATAAGATGATTCTCTGCCCCCACTCTACCACTCTCCCAGCTTCTTTCCTCTTTTTTTTTTTTTTTTTTTTTAATGAGTTGGAGTCTTGCTCTTTCACCCAGTCTGGAGTGTAGTGGCAGGATCACAGCTCACTGCAGCCTTGAACTCCTGGGCTCAAGGGATCCTCCTGCCTCAGCCTCCCAAGTAGCTGGAACTACAGGCGTTTGGTATGACGCCTAGCTAATATTTTCATTTTTCGTAGAGATGCGGTCTTGCTGTGTTGCCTAGGTTGGTCTCTAACTCCTGGCCTCCAGCAG...
[ "visual perception", "negative regulation of lymphangiogenesis", "thyroid hormone transport", "cellular response to hydrogen peroxide", "positive regulation of gastrulation", "intermembrane phospholipid transfer", "amine transport", "renal D-glucose absorption", "positive regulation of vascular asso...
Which one of the following Gene Ontology biological-process terms is a supported annotation for the gene contained within this DNA span? <dna_placeholder> Choose the best answer: A) visual perception B) negative regulation of lymphangiogenesis C) thyroid hormone transport D) cellular response to hydrogen peroxide E) po...
[ 0 ]
[ "A" ]
[ "visual perception" ]
{ "organism": "Homo sapiens", "taxonomic_group": "vertebrate_mammalian", "assembly_accession": "GCF_000001405.40", "sequence_accession": "NC_000019.10", "gene_id": "gene-OPA3", "strand": -1, "chunk_index": 45, "gene_start_chunk": 527426, "gene_end_chunk": 584802, "dna_start_chunk": 527328, "dna_en...
{ "canonical_answer": "visual perception", "entities": [ "visual perception" ], "go_ids": [ "GO:0007601" ], "localization_labels": [], "sources": [ "goa_uniprot_all_gaf_20260410" ], "quality": 1, "source_task_id": "synthesis.aspect.biological_process" }
{ "method": "atomic_go_v5", "source_gene_ids": [ "gene-Ngp", "gene-SLC16A10", "gene-Cdk1", "gene-Lhx1", "gene-VAT1", "gene-SLC22A3", "gene-Hnf1a", "gene-Chrm3", "gene-EDN2" ], "source_sample_ids": [ "17c685dd05d07d08d7ac5ce78b27c1c9878dc9e04e6b0138594c8388e832b48f", "...
{ "dataset_name": "GeneAnnotationBench", "source_sample_id": "GCF_000001405.40::gene-OPA3::NC_000019.10:45::synthesis.aspect.biological_process", "source_task_id": "synthesis.aspect.biological_process", "benchmark_task_id": "bench.aspect_biological_process_mc", "source_split": "eval", "row_seed": "153864942...
{ "schema_version": 1, "policy_version": "atomic-go-v5.0", "selection_seed": 20260804, "gold_go_id": "GO:0007601", "gold_go_name": "visual perception", "gold_go_depth": 6, "gold_prevalence": 0.00036946421651608166, "gold_information_content": 7.903456665253203, "gold_evidence_code": "IMP", "gold_sou...
{"biological_review": ["isoform_geometry_requires_choice_review"], "dna_changed": false, "holds": [], "policy": "gab-conditional-targeting-v1", "source_record_sha256": "c3f6e89fa1e744f1a1c4632c09fb3944b347c35ef35917bd0c18e6b58c1469a8", "targeting": {"mode": "none", "text": ""}}
GCF_000001405.40::gene-SRP72::NC_000004.12:56::synthesis.aspect.biological_process
biological_process
single_choice
Which one of the following Gene Ontology biological-process terms is a supported annotation for the gene contained within this genomic window?
CACTCCCAACTCGGCGACGCTTAGCAATCATCGACTTCCTCCTCCTCTTGGCTGCCTCGGAGATCCTGTTCCGGGGCAGAGGTCTCCCCGCCCCGCCCCTCGTCTCCTCCAAGATGGCGAGCGGCGGCAGCGGGGGGGTGTCAGTACCTGCGCTGTGGAGTGAAGTGAACCGGTATGGCCAGAACGGCGACTTCACGCGCGCTCTCAAGACCGTCAATAAGAGTAAGTGTCGGGGTGGGCACTGGGGCGGGCCCAGGCCGGCTGGAGGATGCGGCCTGTTTCCGGCGCGCGAGACCACCTCGCGCGGGAGGCACGGGAAG...
[ "thymidine metabolic process", "synapse assembly involved in innervation", "dolichol phosphate mannose biosynthetic process", "excitatory synapse pruning", "mitochondrial unfolded protein response", "SRP-dependent cotranslational protein targeting to membrane, signal sequence recognition", "B cell media...
Which one of the following Gene Ontology biological-process terms is a supported annotation for the gene contained within this genomic window? <dna_placeholder> Choose the best answer: A) thymidine metabolic process B) synapse assembly involved in innervation C) dolichol phosphate mannose biosynthetic process D) excita...
[ 5 ]
[ "F" ]
[ "SRP-dependent cotranslational protein targeting to membrane, signal sequence recognition" ]
{ "organism": "Homo sapiens", "taxonomic_group": "vertebrate_mammalian", "assembly_accession": "GCF_000001405.40", "sequence_accession": "NC_000004.12", "gene_id": "gene-SRP72", "strand": 1, "chunk_index": 56, "gene_start_chunk": 467616, "gene_end_chunk": 503681, "dna_start_chunk": 467522, "dna_en...
{ "canonical_answer": "SRP-dependent cotranslational protein targeting to membrane, signal sequence recognition", "entities": [ "SRP-dependent cotranslational protein targeting to membrane, signal sequence recognition" ], "go_ids": [ "GO:0006617" ], "localization_labels": [], "sources": [ "goa...
{ "method": "atomic_go_v5", "source_gene_ids": [ "gene-Tk2", "gene-Lrit3", "gene-DPM3", "gene-Trem2", "gene-ABCB10", "gene-CD70", "gene-Kcna5", "gene-PDXP", "gene-Tnni3" ], "source_sample_ids": [ "9404cbb9b7c1233d786ef160b7c5906b1a00759e3952a394225454a5dfec9ac5", "24d...
{ "dataset_name": "GeneAnnotationBench", "source_sample_id": "GCF_000001405.40::gene-SRP72::NC_000004.12:56::synthesis.aspect.biological_process", "source_task_id": "synthesis.aspect.biological_process", "benchmark_task_id": "bench.aspect_biological_process_mc", "source_split": "eval", "row_seed": "94355538...
{ "schema_version": 1, "policy_version": "atomic-go-v5.0", "selection_seed": 20260804, "gold_go_id": "GO:0006617", "gold_go_name": "SRP-dependent cotranslational protein targeting to membrane, signal sequence recognition", "gold_go_depth": 6, "gold_prevalence": 0.00004134527135893618, "gold_information_...
{"biological_review": ["isoform_geometry_requires_choice_review"], "dna_changed": false, "holds": [], "policy": "gab-conditional-targeting-v1", "source_record_sha256": "61c84e427aef3828631f310653f25e793a795bb1447730ba5bc2d6888c247f58", "targeting": {"mode": "none", "text": ""}}
GCF_000001405.40::gene-ANO7::NC_000002.12:241::synthesis.aspect.biological_process
biological_process
single_choice
"Which one of the following Gene Ontology biological-process terms is a supported annotation for the(...TRUNCATED)
"AATAAGGTTCCCAGGAGGACGGCAGCATTTGCATGCGGGTGTGTGAAAGCCCAGCCAAGTGGGCTGCCTCTGTCTGCAGCTCTGAGCCCCACACCCCCT(...TRUNCATED)
["dichotomous subdivision of terminal units involved in lung branching","negative regulation of lipo(...TRUNCATED)
"Which one of the following Gene Ontology biological-process terms is a supported annotation for the(...TRUNCATED)
[ 2 ]
[ "C" ]
[ "calcium activated phospholipid scrambling" ]
{"organism":"Homo sapiens","taxonomic_group":"vertebrate_mammalian","assembly_accession":"GCF_000001(...TRUNCATED)
{"canonical_answer":"calcium activated phospholipid scrambling","entities":["calcium activated phosp(...TRUNCATED)
{"method":"atomic_go_v5","source_gene_ids":["gene-Ctsh","gene-Angptl4","gene-LOC101404829","gene-SRP(...TRUNCATED)
{"dataset_name":"GeneAnnotationBench","source_sample_id":"GCF_000001405.40::gene-ANO7::NC_000002.12:(...TRUNCATED)
{"schema_version":1,"policy_version":"atomic-go-v5.0","selection_seed":20260804,"gold_go_id":"GO:006(...TRUNCATED)
"{\"biological_review\": [\"isoform_geometry_requires_choice_review\"], \"dna_changed\": false, \"ho(...TRUNCATED)
GCF_000001405.40::gene-GAL3ST2::NC_000002.12:241::synthesis.aspect.domains
domains
single_choice
What structural domains are described for the protein encoded within this genomic region?
"TGTTGGAGCCCACCAAGCCCGTCCCTTGTGGGGACAGCTCCCTCCAAGCCCCCCGGGGGGATGGACACTGAGGGTGGGGCAGGGGAGCAGGGGTAGGCC(...TRUNCATED)
["It contains UbiE/COQ5 methyltransferase, S-adenosyl-L-methionine-dependent methyltransferase super(...TRUNCATED)
"What structural domains are described for the protein encoded within this genomic region?\n<dna_pla(...TRUNCATED)
[ 6 ]
[ "G" ]
["It contains Galactose-3-O-sulfotransferase and P-loop containing nucleoside triphosphate hydrolase(...TRUNCATED)
{"organism":"Homo sapiens","taxonomic_group":"vertebrate_mammalian","assembly_accession":"GCF_000001(...TRUNCATED)
{"canonical_answer":"It contains Galactose-3-O-sulfotransferase and P-loop containing nucleoside tri(...TRUNCATED)
{"method":"random_trainval","source_gene_ids":["BMZ51_RS13120","I4I82_RS33045","LOC107265610","AYR66(...TRUNCATED)
{"dataset_name":"GeneAnnotationBench","source_sample_id":"GCF_000001405.40::gene-GAL3ST2::NC_000002.(...TRUNCATED)
null
"{\"biological_review\": [], \"dna_changed\": false, \"holds\": [], \"policy\": \"gab-conditional-ta(...TRUNCATED)
GCF_000001405.40::gene-ATP6V1C1::NC_000008.11:103::synthesis.aspect.family
family
single_choice
Which family classification is associated with the protein encoded within this genomic window?
"AGCGGCTGCGGCTGCGCAGTGGCGCGGGCGTAGGCGGAGCAGCGCGCGCCGCGGTCAGCTGACTGCTGGGCTGGCACGTGACTTGTTCTGTGTTCGCTT(...TRUNCATED)
["It is associated with the gAF-like domain superfamily family.","It is associated with the hAD hydr(...TRUNCATED)
"Which family classification is associated with the protein encoded within this genomic window?\n<dn(...TRUNCATED)
[ 9 ]
[ "J" ]
[ "It is associated with the vacuolar ATP synthase subunit C superfamily family." ]
{"organism":"Homo sapiens","taxonomic_group":"vertebrate_mammalian","assembly_accession":"GCF_000001(...TRUNCATED)
{"canonical_answer":"It is associated with the vacuolar ATP synthase subunit C superfamily family.",(...TRUNCATED)
{"method":"random_trainval","source_gene_ids":["MAE01_RS12430","POL72_RS48700","ENDOG","OIE73_RS1380(...TRUNCATED)
{"dataset_name":"GeneAnnotationBench","source_sample_id":"GCF_000001405.40::gene-ATP6V1C1::NC_000008(...TRUNCATED)
null
"{\"biological_review\": [], \"dna_changed\": false, \"holds\": [], \"policy\": \"gab-conditional-ta(...TRUNCATED)
GCF_000001405.40::gene-PSMB8::NC_000006.12:32::synthesis.aspect.family
family
single_choice
Which protein family or class is encoded within this sequence window?
"ATTTCTTAGATAGGGGTCTGAGCCCCTTCTCCAGAAAAAGCGTCTCTTTACTTTCTGCCCCACCCAACAACCACAGGCCCAACCCCATTCAGCCACAAG(...TRUNCATED)
["It is associated with the jann4075-like superfamily family.","It belongs to the peptidase T1B fami(...TRUNCATED)
"Which protein family or class is encoded within this sequence window?\n<dna_placeholder>\nChoose th(...TRUNCATED)
[ 1 ]
[ "B" ]
[ "It belongs to the peptidase T1B family." ]
{"organism":"Homo sapiens","taxonomic_group":"vertebrate_mammalian","assembly_accession":"GCF_000001(...TRUNCATED)
{"canonical_answer":"It belongs to the peptidase T1B family.","entities":["Belongs to the peptidase (...TRUNCATED)
{"method":"random_trainval","source_gene_ids":["EV140_RS04585","C7389_RS10430","EP380_RS05585","tbpl(...TRUNCATED)
{"dataset_name":"GeneAnnotationBench","source_sample_id":"GCF_000001405.40::gene-PSMB8::NC_000006.12(...TRUNCATED)
null
"{\"biological_review\": [\"isoform_geometry_requires_choice_review\"], \"dna_changed\": false, \"ho(...TRUNCATED)
GCF_000001405.40::gene-LACC1::NC_000013.11:43::synthesis.aspect.family
family
single_choice
Which family classification is associated with the protein encoded within this genomic region?
"CAAATGAACAACAAAAAAACACTAAAACATCTCGAGGGGGCCCCTGACTTGAGCCCTGCATGGGAAAAGCTGCCAAGAGCTAGCATGACAAGGGACTGG(...TRUNCATED)
["It is associated with the fAD/NAD(P)-binding domain superfamily family.","It is associated with th(...TRUNCATED)
"Which family classification is associated with the protein encoded within this genomic region?\n<dn(...TRUNCATED)
[ 9 ]
[ "J" ]
[ "It belongs to the purine nucleoside phosphorylase YfiH/LACC1 family." ]
{"organism":"Homo sapiens","taxonomic_group":"vertebrate_mammalian","assembly_accession":"GCF_000001(...TRUNCATED)
{"canonical_answer":"It belongs to the purine nucleoside phosphorylase YfiH/LACC1 family.","entities(...TRUNCATED)
{"method":"random_trainval","source_gene_ids":["ACFOHJ_RS00805","HRM2_RS00475","HU739_RS09360","BMZ5(...TRUNCATED)
{"dataset_name":"GeneAnnotationBench","source_sample_id":"GCF_000001405.40::gene-LACC1::NC_000013.11(...TRUNCATED)
null
"{\"biological_review\": [\"isoform_geometry_requires_choice_review\"], \"dna_changed\": false, \"ho(...TRUNCATED)
GCF_000001405.40::gene-ERCC1::NC_000019.10:45::synthesis.aspect.family
family
single_choice
Which family classification is associated with the protein encoded within this DNA span?
"TTCAGGCCCCTGCAGACTTTGCCCCAGAATGGTGAGTGGTCTTGTTGACGGAAAAGAGGGTCCCGGTCCAGACCCCAAGAGCGGGTTCTTGAATTTGTC(...TRUNCATED)
["It is associated with the aminomethyltransferase superfamily family.","It is associated with the c(...TRUNCATED)
"Which family classification is associated with the protein encoded within this DNA span?\n<dna_plac(...TRUNCATED)
[ 4 ]
[ "E" ]
[ "It belongs to the ERCC1/RAD10/SWI10 family." ]
{"organism":"Homo sapiens","taxonomic_group":"vertebrate_mammalian","assembly_accession":"GCF_000001(...TRUNCATED)
{"canonical_answer":"It belongs to the ERCC1/RAD10/SWI10 family.","entities":["Belongs to the ERCC1/(...TRUNCATED)
{"method":"random_trainval","source_gene_ids":["BCF89_RS00205","HA48_RS02520","LZ016_RS09760","GALF_(...TRUNCATED)
{"dataset_name":"GeneAnnotationBench","source_sample_id":"GCF_000001405.40::gene-ERCC1::NC_000019.10(...TRUNCATED)
null
"{\"biological_review\": [\"isoform_geometry_requires_choice_review\"], \"dna_changed\": false, \"ho(...TRUNCATED)
GCF_000001405.40::gene-CAPN3::NC_000015.10:42::synthesis.aspect.family
family
single_choice
How is the protein encoded within this region of DNA classified by family?
"CTCTCATACTCACAGTGTTTCTTCACATCTATTTTTAGTTTTCCTGGCTCAAGCATCTTCAGGCCACTGAAACACAACCCTCACTCTCTTTCTCTCTCC(...TRUNCATED)
["It is associated with the metI-like superfamily family.","It is associated with the ta0600-like su(...TRUNCATED)
"How is the protein encoded within this region of DNA classified by family?\n<dna_placeholder>\nChoo(...TRUNCATED)
[ 5 ]
[ "F" ]
[ "It belongs to the peptidase C2 family." ]
{"organism":"Homo sapiens","taxonomic_group":"vertebrate_mammalian","assembly_accession":"GCF_000001(...TRUNCATED)
{"canonical_answer":"It belongs to the peptidase C2 family.","entities":["Belongs to the peptidase C(...TRUNCATED)
{"method":"random_trainval","source_gene_ids":["N24_RS16130","CQ14_RS18990","FIB18_RS09545","ACFOWZ_(...TRUNCATED)
{"dataset_name":"GeneAnnotationBench","source_sample_id":"GCF_000001405.40::gene-CAPN3::NC_000015.10(...TRUNCATED)
null
"{\"biological_review\": [\"replacement_pending_fresh_biological_review\", \"isoform_geometry_requir(...TRUNCATED)
End of preview. Expand in Data Studio

GeneAnnotationBench

Publishing Metadata

  • Domain: Science (biology / genomics).
  • Modality: Text and tabular; DNA is represented as nucleotide sequences.
  • Capability breakdown: Seven gene-annotation aspects, 500 questions each (1/7 of the dataset, approximately 14.29% each): biological_process, domains, family, function, localization, molecular_activity, pathways.
  • Source: Blend of public biological reference resources, with automated processing and manual review.
  • Size bin: 1K<n<10K (3,500 questions).
  • License: CC BY 4.0.

Dataset Description

GeneAnnotationBench is a frozen evaluation set for testing whether models can interpret raw DNA sequences well enough to answer questions about gene biology. It recasts gene annotation as DNA-conditioned question answering: much as a visual question-answering benchmark pairs an image with a question, GeneAnnotationBench pairs a DNA sequence with a natural-language multiple-choice question.

The dataset contains 3,500 questions across seven gene-annotation aspects. Each question provides a DNA sequence and ten candidate answers. Six aspects use single-choice questions; subcellular localization uses multi-select questions because a gene product may occur in more than one location.

The benchmark is intended to measure sequence-grounded biological understanding, not recall from gene names or annotation databases. Rich provenance is included for audit and error analysis, but is kept separate from the model-input fields.

This dataset is ready for commercial or non-commercial use.

Coverage-repaired revision without targeting hints (2026-10-22)

test.parquet is the qualified unambiguous-v6 revision: 3,500 questions, seven aspects with 500 questions each, and nested 350/700 panels with 50/100 questions per aspect. There are zero model-facing target-location hints in the 350, 700, and full 3,500-question sets. Target localization remains part of the benchmark task.

Independent GPT-6 Astra and Opus 5 reviewers assessed 160 flagged windows, including all 73 windows containing a complete neighboring coding product and 49 flagged questions in the 700 panel. They agreed on five target-ambiguity replacements: two in both nested panels and three corpus-filler questions. Each replacement preserves organism, aspect, and question type and comes from the private pool with complete target CDS coverage and no annotated neighboring CDS overlap. Exact replacement records, including the revised yeast key, received two independent reviews. Existing gene-dense windows are retained when supported neighbor interpretations do not change the offered answer set.

A separate bounded check corrected 11 retained filler localization keys for positively supported offered annotations, including upward GO is_a/part_of ancestry. These key amendments are outside the 350/700 panels and were reviewed as exact proposals. The seven-aspect structure is unchanged.

The 700-question panel inherits the earlier biological review and the current exact replacement review. This is not new individual certification of all other 2,800 rows. Unknown annotations remain evidence limitations, and absence of an annotation is not proof that a biological alternative is impossible. Computational GO assertions and inferred pathway associations are retained as annotation support, rather than experimentally demonstrated function or location. Review scope and source provenance are retained separately from model inputs.

All declared required CDS segments are present, sequence windows fit the 131,072-base DNA budget, and DNA is in forward-reference orientation. The pinned alignment/SFT provenance audit found no remaining target-locus/window matches; this is not a claim about arbitrary pretraining corpora.

Changed prompts, gold labels, and panel members require a fresh benchmark run; pin the Hub commit when reporting scores. This dataset release does not include encoder features, model predictions, or benchmark scores.

Revision lineage (SHA-256 of test.parquet):

  • cabd06d8f8f0e8be5b677a75706e1b919d3d73e179db01a0d1636286c44eb822 — previous coverage-repaired conditional-targeting-v4, preserved in Hub history.
  • bd64157bb8d72a625647b75abc50cf489ae9a78e5d0e84c3f499ced1a79341c5 — this revision, unambiguous-v6 without targeting hints.

Dataset Owner

NVIDIA Corporation

Dataset Creation Date

July 30, 2026

Version

v1 (first public release).

Previous public versions: No previous version.

Content lineage: bench-mc-v5.0 base, revised 2026-10-22 as unambiguous-v6 with coverage repair, reviewed private-pool backfills, supported-key corrections and no target-location hints. Hugging Face commit SHAs version this export. The SHA-256 of test.parquet is bd64157bb8d72a625647b75abc50cf489ae9a78e5d0e84c3f499ced1a79341c5.

License/Terms of Use

GeneAnnotationBench is licensed under the Creative Commons Attribution 4.0 International License (CC BY 4.0). You may share and adapt the dataset for commercial or non-commercial purposes, provided you give appropriate credit, provide a link to the license, and indicate if changes were made.

Copyright © 2026 NVIDIA Corporation. Please credit NVIDIA Corporation for GeneAnnotationBench and retain the upstream source attributions and terms below.

The table below describes the upstream material present in this frozen export. Question counts refer to the source recorded for the gold answer in annotation_provenance.sources.

Upstream resource Material used Question aspects Terms
NCBI RefSeq Genomic FASTA sequence and GFF3 or assembly metadata were used to extract the DNA window and record organism, assembly, sequence, gene, strand, and coordinate provenance. RefSeq is the DNA source for all 3,500 questions; it is not the source of the gold annotation labels in this release. All aspects NCBI states that it places no restrictions on the use or distribution of its molecular data, while noting that original submitters or jurisdictions may assert rights in some submitted material. See the NCBI molecular-data policy.
Gene Ontology and UniProt-GOA UniProt accession-to-GO-term associations, GO aspect codes, GO identifiers, and human-readable GO term names. Biological Process (P) terms form 500 biological_process answers; Molecular Function (F) terms form 500 molecular_activity answers; Cellular Component (C) terms form 349 localization answers. biological_process (500), molecular_activity (500), localization (349) GO data and data products are CC BY 4.0. UniProt applies CC BY 4.0 to copyrightable parts of its databases.
UniProtKB/Swiss-Prot and UniProt identifier mapping Curated FUNCTION, SIMILARITY, and SUBCELLULAR LOCATION comment text supplies the corresponding answer labels. RefSeq-to-UniProt accession mapping also provides the identifier bridge used to join GO, InterPro/Pfam, and Reactome annotations to RefSeq proteins. function (500), family (252), localization (151) UniProt applies CC BY 4.0 to copyrightable parts of its databases and notes that some records may be subject to additional patent or other rights.
InterPro and Pfam InterPro entry identifiers and names plus member-database signature identifiers and names, including Pfam signatures. Combined InterPro/Pfam annotations form all 500 domains answers; Pfam classifications form 248 family answers. domains (500), family (248) InterPro and Pfam downloadable data provided through InterPro are available under CC0 1.0. No InterProScan software is included or used as dataset content.
Reactome The UniProt2Reactome.txt annotation mapping was used to associate UniProt accessions with Reactome stable pathway identifiers and pathway names. These names form all 500 pathways answers. No Reactome software, database dump, pathway illustration, icon, or other artwork is included. pathways (500) Reactome database data and files derived from those data are CC0. The separate terms for Reactome software and visual materials do not apply to the material used here.

The source-derived annotations are used to construct both gold choices and same-aspect distractor pools. Every released row records its gold-label source, while distractor_provenance identifies the source samples used to construct the distractors. Rows minimally repaired during the individual review record an individual_mc_review source marker in annotation_provenance.sources; the upstream-mix counts above refer to the underlying bench-mc-v5.0 export.

Users are responsible for complying with CC BY 4.0 and all applicable upstream terms.

Intended Usage

GeneAnnotationBench is intended for held-out evaluation of models that predict gene annotations from raw DNA. It can be used to:

  • compare DNA-native, multimodal, and text-only models;
  • measure performance across distinct annotation aspects;
  • study how performance changes with sequence length, organism, or annotation source; and
  • audit model errors using the included provenance after inference.

The benchmark should not be used as training data, as a source of clinical conclusions, or as a substitute for experimental validation or expert biological review. Because the release contains gold labels and detailed provenance, it is suitable for open-book local evaluation but not for keeping test answers secret from dataset consumers.

Dataset Characterization

Data Collection Method

Hybrid: Automated, Manually-Collected. DNA windows were derived from reference assemblies, while gene annotations and supporting identifiers were assembled from curated biological resources. The release pipeline automatically normalized records, generated the multiple-choice representation, enforced public/private leakage boundaries, and selected matched evaluation examples.

The public benchmark was selected from a larger candidate pool. Candidates passed structural and provenance checks, automated biological review, and blinded manual adjudication.

Labeling Method

Hybrid: Automated, Manually-Labeled. Gold answers originate from curated source annotations. Candidate questions, answer choices, and provenance were checked for structural consistency and biological support. Cases flagged as contradictory, defective, ambiguous, or insufficiently grounded were reviewed or removed according to the release audit process.

Dataset Format

The dataset is distributed as one Parquet file, test.parquet, containing a single test split. Each row is one DNA-conditioned multiple-choice question. The file has 16 top-level columns, including structured provenance and a JSON geometry/review audit.

Two JSON subset files record the individually reviewed evaluation panels. subsets/700.json lists 700 reviewed question ids (100 per aspect) and subsets/350.json a nested 350 (50 per aspect). The ids key is authoritative; these are review panels, not fresh seeded random draws. The subsets select rows of the test split by id and are not separate dataset splits.

Columns

Column Type Description
id string Frozen source sample ID
aspect string Gene-annotation aspect
question_type string single_choice or multi_select
question string Natural-language question
dna_sequence string Uppercase ACGTN DNA sequence; model input
choices list of string Ten ordered answer choices
prompt_template string Complete text template containing one DNA placeholder; model input
answer_indices list of int8 Zero-based correct-choice indices; gold label
answer_letters list of string Corresponding A–J labels; gold label
answer_texts list of string Corresponding choice texts; gold label
dna_provenance struct Organism, assembly, gene, coordinates, clipping, length, and checksum metadata
annotation_provenance struct Canonical annotation, entities, ontology or location labels, source task, quality, and sources
distractor_provenance struct Distractor-generation method and source identifiers or splits
release_provenance struct Stable source IDs, task IDs, source split, seed, and frozen row ordering
go_provenance nullable struct Atomic-GO gold term, full truth set, choice ontology metadata, source annotations, evidence, prevalence, release hashes, policy, and seed for the refreshed biological-process and molecular-function questions
geometry_provenance JSON string Target CDS geometry, coverage and reviewed repair/admission audit; provenance only

Coordinates in dna_provenance are zero-based and half-open. Fields ending in _chunk are relative to a 1,000,000-bp source chunk; gene_start, gene_end, dna_start, and dna_end are the corresponding global coordinates. DNA is supplied in forward-reference orientation. The strand field is provenance and does not imply that consumers should reverse-complement the sequence.

Dataset Quantification

Property Value
Record count (questions) 3,500
Annotation aspects 7
Questions per aspect 500
Single-choice questions 3,000
Multi-select localization questions 500
Multi-select questions with more than one correct option 149
Choices per question 10
Unique DNA sequences 3,373
Repeated sequence occurrences across aspects 127
Total DNA bases represented 13,097,046
DNA sequence length 262–123,884 bp (median 2,459.5 bp)
Feature count (top-level columns) 16
Total storage size (Parquet file) 13.25 MB

There are 500 questions for each aspect:

  • biological_process
  • domains
  • family
  • function
  • localization
  • molecular_activity
  • pathways

Repeated DNA sequences occur when the same sequence supports questions about different annotation aspects; these are distinct questions, not duplicate rows. Current windows pass the declared CDS coverage and DNA budget checks; original flank-augmentation metadata is retained alongside the geometry repair provenance.

Example question

The following is the first row’s current model-facing template. DNA is supplied separately through dna_sequence; gold labels and provenance are omitted.

Which one of the following Gene Ontology biological-process terms is a supported annotation for the gene contained within this stretch of DNA?
<dna_placeholder>
Choose the best answer:
A) regulation of translation in response to endoplasmic reticulum stress
B) copper ion transport
C) dTMP salvage
D) negative regulation of adenylate cyclase-activating adrenergic receptor signaling pathway
E) positive regulation of natural killer cell chemotaxis
F) mitochondrial tRNA wobble position uridine thiolation
G) fast-twitch skeletal muscle fiber contraction
H) regulation of transcription by RNA polymerase II
I) nuclear polyadenylation-dependent rRNA catabolic process
J) protein K33-linked deubiquitination

Answer with a single letter.

Evaluation

Load a Pinned Revision

Install datasets, copy the immutable commit SHA from the Hub's Files and versions page, and pin it as REVISION:

from datasets import load_dataset

REVISION = "<commit-sha>"
test = load_dataset(
    "nvidia/GeneAnnotationBench",
    split="test",
    revision=REVISION,
)

Model-Input Contract

Only prompt_template and dna_sequence are model inputs. Fields named answer_* and *_provenance are labels or provenance and must never be appended to, rendered into, or otherwise exposed through the evaluation prompt. Provenance supports traceability and post-inference analysis only; it may reveal the answer directly or indirectly.

Every template contains exactly one <dna_placeholder>. DNA-native models should retain that placeholder in the text prompt and pass the sequence through their DNA modality. Text-only baselines should replace the placeholder with the uppercase ACGTN sequence.

row = test[0]
assert row["prompt_template"].count("<dna_placeholder>") == 1

# DNA-native model
prompt = row["prompt_template"]
dna = row["dna_sequence"]
# prediction = model.generate(prompt=prompt, dna_sequence=dna)

# Text-only baseline
text_prompt = row["prompt_template"].replace(
    "<dna_placeholder>", row["dna_sequence"]
)
# prediction = model.generate(text_prompt)

# Never add row["answer_indices"], other answer_* fields, or *_provenance fields.

Models should return one option letter for single_choice questions and a comma-separated set of option letters for multi_select questions.

Final-answer format and optional reasoning

For a boxed-answer evaluation protocol, replace only the final plain-letter instruction in prompt_template with:

Respond with exactly one final answer in \boxed{...}, containing the option
letter or letters. For multiple letters, use alphabetical order separated by
comma and space, for example \boxed{A, F}.

Preserve the question, choices, their order, DNA placeholder and gold answer_indices. This instruction does not require a visible reasoning trace. It works with models that hide reasoning or return it separately as reasoning_content. Score only the final answer in the assistant content; never concatenate a separate reasoning field into the answer being scored. For a native model returning an inline trace, separate that trace from its final answer before applying the same answer parser.

import re

def parse_boxed_answer(final_answer):
    """Parse final answer content independently of reasoning availability."""
    match = re.fullmatch(r"\s*\\boxed\{([A-J](?:, [A-J])*)\}\s*", final_answer)
    if match is None:
        return []
    letters = match.group(1).split(", ")
    if letters != sorted(set(letters)):
        return []
    return ["ABCDEFGHIJ".index(letter) for letter in letters]

Use the same single-choice exact match and localization option-set micro-F1 under this protocol. Report the prompt version, answer parser, token budget and malformed-answer rate. Keep historical plain-letter scores identified as such; their answer protocol differs. The relaxed scoring implementation below accepts explicit answers accompanied by explanations. Never concatenate separate reasoning into scored content.

Trace-envelope checks, such as the native Lightning newline and closing-tag contract, belong to training diagnostics and GRPO formatting rewards. Report these separately from benchmark answer scores. Missing, hidden or separately returned reasoning must not invalidate an otherwise correct benchmark answer. This documentation does not modify the frozen test.parquet or its SHA-256.

Scoring: relaxed explicit-answer extraction

The current relaxed scorer accepts explanatory assistant text when it contains an explicit, unambiguous answer. It uses only the final assistant content and presented choices; separate thinking, reasoning summaries, gold labels and provenance are never inputs to answer extraction. This is deterministic parsing, not an LLM judge or semantic answer matching.

The parser first retains answers accepted by the original strict parser. If that fails, it tries consistent boxed selections, answer-labelled lines, then an explicit leading, bolded or terminal selection. Full choice text is accepted only when it matches exactly one presented choice, ignoring case, presentation markup and a trailing period. It does not harvest option letters from prose or accept a copied choice list as an answer. At the selected extraction priority, conflicting selections, duplicates, alternatives such as “A or C”, and multiple letters for a single-choice question score as an empty prediction.

Examples:

Assistant content Question type Parsed selection
Final answer: C. Because A and B are distractors. single choice C
**C) siRNA processing** single choice C
Answer: A and C — both compartments. multi-select A, C
Analysis mentions A and B. followed by The answer is C. single choice C
Answer: A or C either empty
Answer: A, A either empty
Answer: A, C single choice empty
Answer: A followed by Answer: C either empty

For the six single-choice aspects, report exact-match accuracy. For localization, aggregate true positives, false positives and false negatives across questions and report option-set micro-F1, 2 TP / (2 TP + FP + FN). The headline score is the unweighted mean of these seven aspect metrics, not pooled question accuracy. An empty selection is incorrect for single-choice questions and adds all gold localization options to false negatives.

The API completion status must be normalized to finish_reason="stop" for a completed answer. Any other status, including token-limit termination, scores as empty even if an early answer appears in the response. Missing responses make a run incomplete; do not silently omit them or report a partial run as a complete benchmark. Transport failures may be retried with the same policy; completed or length-capped answers are not resampled to improve the score.

The implementation below is the current campaign parser and scoring policy. It supports both the full test split and a published reviewed panel, provided all seven aspects are represented. Select the 350-question panel by the IDs in subsets/350.json; it contains 50 questions per aspect. Use the subset files from the same pinned Hub commit as the Parquet; replacement IDs differ from earlier reviewed panels.

import re
from collections import Counter

def final_options(content: str, multi_select: bool) -> set[str]:
    """Parse one boxed final answer or an answer-only A–J list, never reasoning.

    Extra boxes, malformed lists, duplicate options and missing answers score as
    empty predictions. A single-choice response containing multiple options also
    scores empty. Parsing is fixed independently of gold labels.
    """
    boxes = re.findall(r'\\boxed\{([^{}]*)\}', content)
    if len(boxes) > 1:
        return set()
    answer = boxes[0] if boxes else content.strip()
    if not re.fullmatch(r'\s*[A-J](?:\s*,\s*[A-J])*\s*', answer):
        return set()
    letters = [value.strip() for value in answer.split(',')]
    if len(set(letters)) != len(letters) or (not multi_select and len(letters) != 1):
        return set()
    return set(letters)


def plain(text):
    """Remove presentation markup while retaining option letters and prose."""
    return re.sub(r'[*`]', '', text).strip().lstrip('#').strip()


def selection(text, multi_select):
    """Read an explicit initial letter list without harvesting prose letters.

Accept commas, semicolons, slashes, ampersands and 'and'. Reject alternatives
expressed with 'or', duplicate letters and multiple single-choice selections.
The suffix may contain the answer text or an explanation.
"""
    text = plain(text)
    match = re.match(r'^\(?([A-J])\b((?:\s*(?:,|;|/|&|\band\b)\s*[A-J]\b)*)', text)
    if not match:
        return None
    suffix = text[match.end():]
    if re.match(r'^\s*(?:or\b|(?:,|;|/|&)\s*(?:or\b|[A-J]\b))', suffix):
        return None
    # A letter must be separated from prose; an initial 'A protein ...' is not
    # an answer selection. Parentheses, punctuation or a line break are explicit.
    if suffix and not re.match(r'^\s*(?:[().,:;\-–—]|$)', suffix):
        return None
    letters = [match.group(1), *re.findall(r'\b[A-J]\b', match.group(2))]
    if len(set(letters)) != len(letters) or (not multi_select and len(letters) != 1):
        return None
    return set(letters)


def explicit_answer(content, multi_select, choices, original_parser):
    """Return a selection plus extraction evidence, or leave ambiguity ungraded.

Priority is original valid answer, consistent boxes, explicit answer-labelled
lines, then a leading selection. Named-choice matching requires a complete
choice text and uniqueness. Copied choice lists are not treated as selections.
"""
    original = original_parser(content, multi_select)
    if original:
        return original, 'original_valid'
    boxes = re.findall(r'\\boxed\{([^{}]*)\}', content)
    if boxes:
        selected = [selection(box, multi_select) for box in boxes]
        if all(selected) and len({tuple(sorted(s)) for s in selected}) == 1:
            return selected[0], 'explicit_box'
        return set(), 'ambiguous_or_invalid_box'
    raw_lines = [line.strip() for line in content.splitlines() if plain(line)]
    lines = [plain(line) for line in raw_lines]
    if not lines:
        return set(), 'empty'
    labelled = []
    for index, line in enumerate(lines):
        if re.fullmatch(r'(?:final\s+|correct\s+)?answer\s*[:=]?', line, re.I):
            if index+1 < len(lines):
                labelled.append(lines[index+1])
            continue
        match = re.match(r'^(?:(?:therefore|thus|so)[,:]?\s*)?(?:the\s+)?(?:final\s+|correct\s+)?answer(?:\s+is)?\s*[:=\-]?\s+(.+)$', line, re.I)
        if not match:
            match = re.match(r'^(?:final\s+|correct\s+)?answer\s*[:=]\s*(.+)$', line, re.I)
        if match:
            labelled.append(match.group(1))
    candidates = labelled or [lines[0]]
    if not labelled and selection(lines[0], multi_select) is None:
        # A separately bolded answer line is an explicit selection even when it
        # follows an analysis paragraph. Multiple different selections conflict.
        bold_selections = [line for raw, line in zip(raw_lines, lines)
                           if raw.startswith('**') and selection(line, multi_select)]
        if bold_selections:
            candidates = bold_selections
        elif selection(lines[-1], multi_select):
            # Some models put an unformatted final selection after explanatory
            # paragraphs. A terminal selection must not merely end a choice list.
            option_lines = [line for line in lines if re.match(r'^[A-J][).]\s', line)]
            if len(option_lines) <= 1 or (re.match(r'^[A-J][).]\s*Correct\b', lines[-1], re.I)
                                         and all(re.match(r'^[A-J][).]\s*Incorrect\b', line, re.I)
                                                 for line in option_lines[:-1])):
                candidates = [lines[-1]]
    parsed = []
    for candidate in candidates:
        result = selection(candidate, multi_select)
        if result is None:
            named = [chr(65+i) for i, choice in enumerate(choices)
                     if plain(candidate).casefold().rstrip('.') == plain(choice).casefold().rstrip('.')]
            result = set(named) if len(named) == 1 else None
        parsed.append(result)
    if not all(parsed) or len({tuple(sorted(s)) for s in parsed}) != 1:
        return set(), 'unresolved_or_conflicting_answer'
    if not labelled and candidates == [lines[0]] and sum(bool(re.match(r'^[A-J][).]\s', line)) for line in lines) > 1:
        return set(), 'unselected_choice_list'
    return parsed[0], 'labelled_answer' if labelled else 'leading_answer'


def score(rows, predictions):
    """Score every row by ID using explicit answers and completed responses.

    predictions[id] must contain content and a normalized finish_reason.
    Missing IDs raise KeyError rather than disappearing from the denominator.
    """
    counts, correct = Counter(), Counter()
    tp = fp = fn = invalid = nonstop = 0
    extraction = []
    for row in rows:
        prediction = predictions[row["id"]]
        selected, rule = explicit_answer(
            prediction["content"], row["aspect"] == "localization",
            row["choices"], final_options,
        )
        if prediction["finish_reason"] != "stop":
            selected, rule = set(), "nonstop_excluded"
            nonstop += 1
        gold = set(row["answer_letters"])
        aspect = row["aspect"]
        counts[aspect] += 1
        invalid += not selected
        if aspect == "localization":
            tp += len(selected & gold)
            fp += len(selected - gold)
            fn += len(gold - selected)
        else:
            correct[aspect] += selected == gold
        extraction.append({"id": row["id"], "selected": sorted(selected), "rule": rule})
    expected = {"biological_process", "domains", "family", "function",
                "localization", "molecular_activity", "pathways"}
    if set(counts) != expected:
        raise ValueError("All seven aspects are required for comparison")
    metrics = {aspect: correct[aspect] / counts[aspect]
               for aspect in sorted(expected - {"localization"})}
    metrics["localization"] = 2 * tp / (2 * tp + fp + fn)
    return {
        "metrics": metrics,
        "mean_seven_aspects": sum(metrics.values()) / 7,
        "counts": dict(counts),
        "invalid_answers": invalid,
        "length_or_other_nonstop_answers": nonstop,
        "extraction": extraction,
    }

Fix the parser before evaluation and use it identically for every model. Record the dataset revision, selected subset, complete prompt, output ceiling, reasoning settings, provider/model identity, finish-status normalization, parser version, and invalid/capped counts. Keep historical strict scores labelled separately; relaxed scoring changes answer extraction, not the underlying metric or labels.

Parser source SHA-256 (the plain, selection, and explicit_answer module): fbbf35373d71a8f67df5458fb10b2370e9ba3e589eec2adcec3c86cd4a98351a.

References

Ethical Considerations

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. Developers should work with their internal developer teams to ensure this dataset meets requirements for the relevant industry and use case and addresses unforeseen product misuse.

GeneAnnotationBench evaluates computational predictions from genomic sequence. Its annotations may be incomplete, uncertain, or context-dependent, and benchmark performance does not establish biological mechanism, clinical validity, or safety. Users should account for organism and source coverage, possible annotation bias, and the risk of overgeneralizing from a fixed multiple-choice benchmark.

Please report quality, risk, security vulnerabilities or NVIDIA AI Concerns here.

Downloads last month
43