Configuration

NT-500M, fine-tuned reference score (Datta et al. 2026)

Evidence concern: excluded from comparisons

Table 1 prints the same frozen value and spread (.5252+-.0262) for NT on both 'Enhancer' and 'Enhancer Types', while the fine-tuned references differ (.548 and .424). One of the two frozen cells is likely a transcription error in the source; both are kept as printed and should not support comparison until the authors clarify.Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, column 'NT', rows 'Enhancer' and 'Enhancer Types'

Table 1 and Table 4 disagree for HyenaDNA Splice Acceptor: Table 1 prints .4752±.0108 (recovery 50), while Table 4 prints 0.7308 ± 0.0033 (validation-selected) and 0.7360 ± 0.0052 (diagnostic grid-best). Five other HyenaDNA cells of Table 1 equal Table 4's grid-best values, but Splice Donor (.7247±.0087) equals the validation-selected value instead. Table 3 and Table 4 also name different grid-best readouts for HyenaDNA Splice All (MLP+Mean against MLP+Attn). Table 1 values are kept as printed.Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, column 'HyenaDNA', rows 'Splice Acceptor' and 'Splice Donor'; Table 4; Table 3, row 'Splice All'

NT-500M, fine-tuned reference score, a configuration of Nucleotide Transformer: 500M parameters.

18 evaluations · 18 results

Overview

An explanatory profile has not yet been reviewed for this record.

Consult the linked sources for architecture or protocol details. Missing evidence is not evidence of a missing capability.

Evaluations and results

18 evaluations · 18 results. Different protocols are not a single leaderboard.

Protocol: Nucleotide Transformer benchmark, reference fine-tuned scores (Datta et al. 2026 Table 1) · Dataset: Nucleotide Transformer benchmark task data (Datta et al. 2026)

Sorted by F1 (NT benchmark Promoter All, reference fine-tuned score) (higher is better). The best value in each column is highlighted. Decimals are rounded for display; each value links to the printed value and its source.

Tested configuration	F1 (NT benchmark Promoter All, reference fine-tuned score)	F1 (NT benchmark Promoter Non-TATA, reference fine-tuned score)	F1 (NT benchmark Promoter TATA, reference fine-tuned score)	F1 (NT benchmark Splice Acceptor, reference fine-tuned score)	F1 (NT benchmark Splice All, reference fine-tuned score)	F1 (NT benchmark Splice Donor, reference fine-tuned score)	MCC (NT benchmark Enhancer Types, reference fine-tuned score)	MCC (NT benchmark Enhancer, reference fine-tuned score)	MCC (NT benchmark H3, reference fine-tuned score)	MCC (NT benchmark H3K14ac, reference fine-tuned score)	MCC (NT benchmark H3K36me3, reference fine-tuned score)	MCC (NT benchmark H3K4me1, reference fine-tuned score)	MCC (NT benchmark H3K4me2, reference fine-tuned score)	MCC (NT benchmark H3K4me3, reference fine-tuned score)	MCC (NT benchmark H3K79me3, reference fine-tuned score)	MCC (NT benchmark H3K9ac, reference fine-tuned score)	MCC (NT benchmark H4, reference fine-tuned score)	MCC (NT benchmark H4ac, reference fine-tuned score)
NT-500M, fine-tuned reference score (Datta et al. 2026)Evaluation	0.976	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported
NT-500M, fine-tuned reference score (Datta et al. 2026)Evaluation	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	0.548	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported
NT-500M, fine-tuned reference score (Datta et al. 2026)Evaluation	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	0.424	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported
NT-500M, fine-tuned reference score (Datta et al. 2026)Evaluation	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	0.784	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported
NT-500M, fine-tuned reference score (Datta et al. 2026)Evaluation	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	0.551	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported
NT-500M, fine-tuned reference score (Datta et al. 2026)Evaluation	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	0.625	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported
NT-500M, fine-tuned reference score (Datta et al. 2026)Evaluation	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	0.55	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported
NT-500M, fine-tuned reference score (Datta et al. 2026)Evaluation	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	0.319	Not reported	Not reported	Not reported	Not reported	Not reported
NT-500M, fine-tuned reference score (Datta et al. 2026)Evaluation	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	0.41	Not reported	Not reported	Not reported	Not reported
NT-500M, fine-tuned reference score (Datta et al. 2026)Evaluation	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	0.672	Not reported	Not reported	Not reported
NT-500M, fine-tuned reference score (Datta et al. 2026)Evaluation	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	0.562	Not reported	Not reported
NT-500M, fine-tuned reference score (Datta et al. 2026)Evaluation	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	0.799	Not reported
NT-500M, fine-tuned reference score (Datta et al. 2026)Evaluation	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	0.495
NT-500M, fine-tuned reference score (Datta et al. 2026)Evaluation	Not reported	0.976	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported
NT-500M, fine-tuned reference score (Datta et al. 2026)Evaluation	Not reported	Not reported	0.966	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported
NT-500M, fine-tuned reference score (Datta et al. 2026)Evaluation	Not reported	Not reported	Not reported	0.981	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported
NT-500M, fine-tuned reference score (Datta et al. 2026)Evaluation	Not reported	Not reported	Not reported	Not reported	0.983	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported
NT-500M, fine-tuned reference score (Datta et al. 2026)Evaluation	Not reported	Not reported	Not reported	Not reported	Not reported	0.985	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported	Not reported

All 18 result rows with coverage, uncertainty and sources

Filter evaluations

Tested model or methodProtocolDatasetEvaluation setupMetricEvidence origin

Applied filters: All linked evaluations

Exact evaluated configurations and original reported results
Tested configuration	Protocol and dataset	Finding	Evidence and details
Configuration: NT-500M, fine-tuned reference score (Datta et al. 2026)	Protocol: Nucleotide Transformer benchmark, reference fine-tuned scores (Datta et al. 2026 Table 1) Dataset: Nucleotide Transformer benchmark task data (Datta et al. 2026)	0.548 MCC (NT benchmark Enhancer, reference fine-tuned score) unitless · higher Uncertainty: ± 0.144 (type not stated). Table 1 does not define the spread on reference fine-tuned scores; Section 3.3 defines only the frozen-probe spread Coverage: Not reported scored / Not reported eligible	Evaluation origin not reported · Source checked Evidence concern: excluded from comparisons Table 1 prints the same frozen value and spread (.5252+-.0262) for NT on both 'Enhancer' and 'Enhancer Types', while the fine-tuned references differ (.548 and .424). One of the two frozen cells is likely a transcription error in the source; both are kept as printed and should not support comparison until the authors clarify.Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, column 'NT', rows 'Enhancer' and 'Enhancer Types' Table 1 and Table 4 disagree for HyenaDNA Splice Acceptor: Table 1 prints .4752±.0108 (recovery 50), while Table 4 prints 0.7308 ± 0.0033 (validation-selected) and 0.7360 ± 0.0052 (diagnostic grid-best). Five other HyenaDNA cells of Table 1 equal Table 4's grid-best values, but Splice Donor (.7247±.0087) equals the validation-selected value instead. Table 3 and Table 4 also name different grid-best readouts for HyenaDNA Splice All (MLP+Mean against MLP+Attn). Table 1 values are kept as printed.Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, column 'HyenaDNA', rows 'Splice Acceptor' and 'Splice Donor'; Table 4; Table 3, row 'Splice All' Methods, coverage and source NT-500M on NT benchmark Enhancer (reference fine-tuned) gfm2-20261011-datta2026-protocol-fine-tuned-reference Aggregation: Not reported Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, row 'Enhancer', column 'NT', FT value
Configuration: NT-500M, fine-tuned reference score (Datta et al. 2026)	Protocol: Nucleotide Transformer benchmark, reference fine-tuned scores (Datta et al. 2026 Table 1) Dataset: Nucleotide Transformer benchmark task data (Datta et al. 2026)	0.424 MCC (NT benchmark Enhancer Types, reference fine-tuned score) unitless · higher Uncertainty: ± 0.132 (type not stated). Table 1 does not define the spread on reference fine-tuned scores; Section 3.3 defines only the frozen-probe spread Coverage: Not reported scored / Not reported eligible	Evaluation origin not reported · Source checked Evidence concern: excluded from comparisons Table 1 prints the same frozen value and spread (.5252+-.0262) for NT on both 'Enhancer' and 'Enhancer Types', while the fine-tuned references differ (.548 and .424). One of the two frozen cells is likely a transcription error in the source; both are kept as printed and should not support comparison until the authors clarify.Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, column 'NT', rows 'Enhancer' and 'Enhancer Types' Table 1 and Table 4 disagree for HyenaDNA Splice Acceptor: Table 1 prints .4752±.0108 (recovery 50), while Table 4 prints 0.7308 ± 0.0033 (validation-selected) and 0.7360 ± 0.0052 (diagnostic grid-best). Five other HyenaDNA cells of Table 1 equal Table 4's grid-best values, but Splice Donor (.7247±.0087) equals the validation-selected value instead. Table 3 and Table 4 also name different grid-best readouts for HyenaDNA Splice All (MLP+Mean against MLP+Attn). Table 1 values are kept as printed.Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, column 'HyenaDNA', rows 'Splice Acceptor' and 'Splice Donor'; Table 4; Table 3, row 'Splice All' Methods, coverage and source NT-500M on NT benchmark Enhancer Types (reference fine-tuned) gfm2-20261011-datta2026-protocol-fine-tuned-reference Aggregation: Not reported Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, row 'Enhancer Types', column 'NT', FT value
Configuration: NT-500M, fine-tuned reference score (Datta et al. 2026)	Protocol: Nucleotide Transformer benchmark, reference fine-tuned scores (Datta et al. 2026 Table 1) Dataset: Nucleotide Transformer benchmark task data (Datta et al. 2026)	0.784 MCC (NT benchmark H3, reference fine-tuned score) unitless · higher Uncertainty: ± 0.047 (type not stated). Table 1 does not define the spread on reference fine-tuned scores; Section 3.3 defines only the frozen-probe spread Coverage: Not reported scored / Not reported eligible	Evaluation origin not reported · Source checked Evidence concern: excluded from comparisons Table 1 prints the same frozen value and spread (.5252+-.0262) for NT on both 'Enhancer' and 'Enhancer Types', while the fine-tuned references differ (.548 and .424). One of the two frozen cells is likely a transcription error in the source; both are kept as printed and should not support comparison until the authors clarify.Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, column 'NT', rows 'Enhancer' and 'Enhancer Types' Table 1 and Table 4 disagree for HyenaDNA Splice Acceptor: Table 1 prints .4752±.0108 (recovery 50), while Table 4 prints 0.7308 ± 0.0033 (validation-selected) and 0.7360 ± 0.0052 (diagnostic grid-best). Five other HyenaDNA cells of Table 1 equal Table 4's grid-best values, but Splice Donor (.7247±.0087) equals the validation-selected value instead. Table 3 and Table 4 also name different grid-best readouts for HyenaDNA Splice All (MLP+Mean against MLP+Attn). Table 1 values are kept as printed.Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, column 'HyenaDNA', rows 'Splice Acceptor' and 'Splice Donor'; Table 4; Table 3, row 'Splice All' Methods, coverage and source NT-500M on NT benchmark H3 (reference fine-tuned) gfm2-20261011-datta2026-protocol-fine-tuned-reference Aggregation: Not reported Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, row 'H3', column 'NT', FT value
Configuration: NT-500M, fine-tuned reference score (Datta et al. 2026)	Protocol: Nucleotide Transformer benchmark, reference fine-tuned scores (Datta et al. 2026 Table 1) Dataset: Nucleotide Transformer benchmark task data (Datta et al. 2026)	0.551 MCC (NT benchmark H3K14ac, reference fine-tuned score) unitless · higher Uncertainty: ± 0.021 (type not stated). Table 1 does not define the spread on reference fine-tuned scores; Section 3.3 defines only the frozen-probe spread Coverage: Not reported scored / Not reported eligible	Evaluation origin not reported · Source checked Evidence concern: excluded from comparisons Table 1 prints the same frozen value and spread (.5252+-.0262) for NT on both 'Enhancer' and 'Enhancer Types', while the fine-tuned references differ (.548 and .424). One of the two frozen cells is likely a transcription error in the source; both are kept as printed and should not support comparison until the authors clarify.Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, column 'NT', rows 'Enhancer' and 'Enhancer Types' Table 1 and Table 4 disagree for HyenaDNA Splice Acceptor: Table 1 prints .4752±.0108 (recovery 50), while Table 4 prints 0.7308 ± 0.0033 (validation-selected) and 0.7360 ± 0.0052 (diagnostic grid-best). Five other HyenaDNA cells of Table 1 equal Table 4's grid-best values, but Splice Donor (.7247±.0087) equals the validation-selected value instead. Table 3 and Table 4 also name different grid-best readouts for HyenaDNA Splice All (MLP+Mean against MLP+Attn). Table 1 values are kept as printed.Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, column 'HyenaDNA', rows 'Splice Acceptor' and 'Splice Donor'; Table 4; Table 3, row 'Splice All' Methods, coverage and source NT-500M on NT benchmark H3K14ac (reference fine-tuned) gfm2-20261011-datta2026-protocol-fine-tuned-reference Aggregation: Not reported Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, row 'H3K14ac', column 'NT', FT value
Configuration: NT-500M, fine-tuned reference score (Datta et al. 2026)	Protocol: Nucleotide Transformer benchmark, reference fine-tuned scores (Datta et al. 2026 Table 1) Dataset: Nucleotide Transformer benchmark task data (Datta et al. 2026)	0.625 MCC (NT benchmark H3K36me3, reference fine-tuned score) unitless · higher Uncertainty: ± 0.013 (type not stated). Table 1 does not define the spread on reference fine-tuned scores; Section 3.3 defines only the frozen-probe spread Coverage: Not reported scored / Not reported eligible	Evaluation origin not reported · Source checked Evidence concern: excluded from comparisons Table 1 prints the same frozen value and spread (.5252+-.0262) for NT on both 'Enhancer' and 'Enhancer Types', while the fine-tuned references differ (.548 and .424). One of the two frozen cells is likely a transcription error in the source; both are kept as printed and should not support comparison until the authors clarify.Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, column 'NT', rows 'Enhancer' and 'Enhancer Types' Table 1 and Table 4 disagree for HyenaDNA Splice Acceptor: Table 1 prints .4752±.0108 (recovery 50), while Table 4 prints 0.7308 ± 0.0033 (validation-selected) and 0.7360 ± 0.0052 (diagnostic grid-best). Five other HyenaDNA cells of Table 1 equal Table 4's grid-best values, but Splice Donor (.7247±.0087) equals the validation-selected value instead. Table 3 and Table 4 also name different grid-best readouts for HyenaDNA Splice All (MLP+Mean against MLP+Attn). Table 1 values are kept as printed.Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, column 'HyenaDNA', rows 'Splice Acceptor' and 'Splice Donor'; Table 4; Table 3, row 'Splice All' Methods, coverage and source NT-500M on NT benchmark H3K36me3 (reference fine-tuned) gfm2-20261011-datta2026-protocol-fine-tuned-reference Aggregation: Not reported Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, row 'H3K36me3', column 'NT', FT value
Configuration: NT-500M, fine-tuned reference score (Datta et al. 2026)	Protocol: Nucleotide Transformer benchmark, reference fine-tuned scores (Datta et al. 2026 Table 1) Dataset: Nucleotide Transformer benchmark task data (Datta et al. 2026)	0.55 MCC (NT benchmark H3K4me1, reference fine-tuned score) unitless · higher Uncertainty: ± 0.021 (type not stated). Table 1 does not define the spread on reference fine-tuned scores; Section 3.3 defines only the frozen-probe spread Coverage: Not reported scored / Not reported eligible	Evaluation origin not reported · Source checked Evidence concern: excluded from comparisons Table 1 prints the same frozen value and spread (.5252+-.0262) for NT on both 'Enhancer' and 'Enhancer Types', while the fine-tuned references differ (.548 and .424). One of the two frozen cells is likely a transcription error in the source; both are kept as printed and should not support comparison until the authors clarify.Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, column 'NT', rows 'Enhancer' and 'Enhancer Types' Table 1 and Table 4 disagree for HyenaDNA Splice Acceptor: Table 1 prints .4752±.0108 (recovery 50), while Table 4 prints 0.7308 ± 0.0033 (validation-selected) and 0.7360 ± 0.0052 (diagnostic grid-best). Five other HyenaDNA cells of Table 1 equal Table 4's grid-best values, but Splice Donor (.7247±.0087) equals the validation-selected value instead. Table 3 and Table 4 also name different grid-best readouts for HyenaDNA Splice All (MLP+Mean against MLP+Attn). Table 1 values are kept as printed.Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, column 'HyenaDNA', rows 'Splice Acceptor' and 'Splice Donor'; Table 4; Table 3, row 'Splice All' Methods, coverage and source NT-500M on NT benchmark H3K4me1 (reference fine-tuned) gfm2-20261011-datta2026-protocol-fine-tuned-reference Aggregation: Not reported Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, row 'H3K4me1', column 'NT', FT value
Configuration: NT-500M, fine-tuned reference score (Datta et al. 2026)	Protocol: Nucleotide Transformer benchmark, reference fine-tuned scores (Datta et al. 2026 Table 1) Dataset: Nucleotide Transformer benchmark task data (Datta et al. 2026)	0.319 MCC (NT benchmark H3K4me2, reference fine-tuned score) unitless · higher Uncertainty: ± 0.045 (type not stated). Table 1 does not define the spread on reference fine-tuned scores; Section 3.3 defines only the frozen-probe spread Coverage: Not reported scored / Not reported eligible	Evaluation origin not reported · Source checked Evidence concern: excluded from comparisons Table 1 prints the same frozen value and spread (.5252+-.0262) for NT on both 'Enhancer' and 'Enhancer Types', while the fine-tuned references differ (.548 and .424). One of the two frozen cells is likely a transcription error in the source; both are kept as printed and should not support comparison until the authors clarify.Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, column 'NT', rows 'Enhancer' and 'Enhancer Types' Table 1 and Table 4 disagree for HyenaDNA Splice Acceptor: Table 1 prints .4752±.0108 (recovery 50), while Table 4 prints 0.7308 ± 0.0033 (validation-selected) and 0.7360 ± 0.0052 (diagnostic grid-best). Five other HyenaDNA cells of Table 1 equal Table 4's grid-best values, but Splice Donor (.7247±.0087) equals the validation-selected value instead. Table 3 and Table 4 also name different grid-best readouts for HyenaDNA Splice All (MLP+Mean against MLP+Attn). Table 1 values are kept as printed.Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, column 'HyenaDNA', rows 'Splice Acceptor' and 'Splice Donor'; Table 4; Table 3, row 'Splice All' Methods, coverage and source NT-500M on NT benchmark H3K4me2 (reference fine-tuned) gfm2-20261011-datta2026-protocol-fine-tuned-reference Aggregation: Not reported Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, row 'H3K4me2', column 'NT', FT value
Configuration: NT-500M, fine-tuned reference score (Datta et al. 2026)	Protocol: Nucleotide Transformer benchmark, reference fine-tuned scores (Datta et al. 2026 Table 1) Dataset: Nucleotide Transformer benchmark task data (Datta et al. 2026)	0.41 MCC (NT benchmark H3K4me3, reference fine-tuned score) unitless · higher Uncertainty: ± 0.033 (type not stated). Table 1 does not define the spread on reference fine-tuned scores; Section 3.3 defines only the frozen-probe spread Coverage: Not reported scored / Not reported eligible	Evaluation origin not reported · Source checked Evidence concern: excluded from comparisons Table 1 prints the same frozen value and spread (.5252+-.0262) for NT on both 'Enhancer' and 'Enhancer Types', while the fine-tuned references differ (.548 and .424). One of the two frozen cells is likely a transcription error in the source; both are kept as printed and should not support comparison until the authors clarify.Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, column 'NT', rows 'Enhancer' and 'Enhancer Types' Table 1 and Table 4 disagree for HyenaDNA Splice Acceptor: Table 1 prints .4752±.0108 (recovery 50), while Table 4 prints 0.7308 ± 0.0033 (validation-selected) and 0.7360 ± 0.0052 (diagnostic grid-best). Five other HyenaDNA cells of Table 1 equal Table 4's grid-best values, but Splice Donor (.7247±.0087) equals the validation-selected value instead. Table 3 and Table 4 also name different grid-best readouts for HyenaDNA Splice All (MLP+Mean against MLP+Attn). Table 1 values are kept as printed.Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, column 'HyenaDNA', rows 'Splice Acceptor' and 'Splice Donor'; Table 4; Table 3, row 'Splice All' Methods, coverage and source NT-500M on NT benchmark H3K4me3 (reference fine-tuned) gfm2-20261011-datta2026-protocol-fine-tuned-reference Aggregation: Not reported Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, row 'H3K4me3', column 'NT', FT value
Configuration: NT-500M, fine-tuned reference score (Datta et al. 2026)	Protocol: Nucleotide Transformer benchmark, reference fine-tuned scores (Datta et al. 2026 Table 1) Dataset: Nucleotide Transformer benchmark task data (Datta et al. 2026)	0.672 MCC (NT benchmark H3K79me3, reference fine-tuned score) unitless · higher Uncertainty: ± 0.048 (type not stated). Table 1 does not define the spread on reference fine-tuned scores; Section 3.3 defines only the frozen-probe spread Coverage: Not reported scored / Not reported eligible	Evaluation origin not reported · Source checked Evidence concern: excluded from comparisons Table 1 prints the same frozen value and spread (.5252+-.0262) for NT on both 'Enhancer' and 'Enhancer Types', while the fine-tuned references differ (.548 and .424). One of the two frozen cells is likely a transcription error in the source; both are kept as printed and should not support comparison until the authors clarify.Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, column 'NT', rows 'Enhancer' and 'Enhancer Types' Table 1 and Table 4 disagree for HyenaDNA Splice Acceptor: Table 1 prints .4752±.0108 (recovery 50), while Table 4 prints 0.7308 ± 0.0033 (validation-selected) and 0.7360 ± 0.0052 (diagnostic grid-best). Five other HyenaDNA cells of Table 1 equal Table 4's grid-best values, but Splice Donor (.7247±.0087) equals the validation-selected value instead. Table 3 and Table 4 also name different grid-best readouts for HyenaDNA Splice All (MLP+Mean against MLP+Attn). Table 1 values are kept as printed.Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, column 'HyenaDNA', rows 'Splice Acceptor' and 'Splice Donor'; Table 4; Table 3, row 'Splice All' Methods, coverage and source NT-500M on NT benchmark H3K79me3 (reference fine-tuned) gfm2-20261011-datta2026-protocol-fine-tuned-reference Aggregation: Not reported Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, row 'H3K79me3', column 'NT', FT value
Configuration: NT-500M, fine-tuned reference score (Datta et al. 2026)	Protocol: Nucleotide Transformer benchmark, reference fine-tuned scores (Datta et al. 2026 Table 1) Dataset: Nucleotide Transformer benchmark task data (Datta et al. 2026)	0.562 MCC (NT benchmark H3K9ac, reference fine-tuned score) unitless · higher Uncertainty: ± 0.040 (type not stated). Table 1 does not define the spread on reference fine-tuned scores; Section 3.3 defines only the frozen-probe spread Coverage: Not reported scored / Not reported eligible	Evaluation origin not reported · Source checked Evidence concern: excluded from comparisons Table 1 prints the same frozen value and spread (.5252+-.0262) for NT on both 'Enhancer' and 'Enhancer Types', while the fine-tuned references differ (.548 and .424). One of the two frozen cells is likely a transcription error in the source; both are kept as printed and should not support comparison until the authors clarify.Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, column 'NT', rows 'Enhancer' and 'Enhancer Types' Table 1 and Table 4 disagree for HyenaDNA Splice Acceptor: Table 1 prints .4752±.0108 (recovery 50), while Table 4 prints 0.7308 ± 0.0033 (validation-selected) and 0.7360 ± 0.0052 (diagnostic grid-best). Five other HyenaDNA cells of Table 1 equal Table 4's grid-best values, but Splice Donor (.7247±.0087) equals the validation-selected value instead. Table 3 and Table 4 also name different grid-best readouts for HyenaDNA Splice All (MLP+Mean against MLP+Attn). Table 1 values are kept as printed.Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, column 'HyenaDNA', rows 'Splice Acceptor' and 'Splice Donor'; Table 4; Table 3, row 'Splice All' Methods, coverage and source NT-500M on NT benchmark H3K9ac (reference fine-tuned) gfm2-20261011-datta2026-protocol-fine-tuned-reference Aggregation: Not reported Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, row 'H3K9ac', column 'NT', FT value
Configuration: NT-500M, fine-tuned reference score (Datta et al. 2026)	Protocol: Nucleotide Transformer benchmark, reference fine-tuned scores (Datta et al. 2026 Table 1) Dataset: Nucleotide Transformer benchmark task data (Datta et al. 2026)	0.799 MCC (NT benchmark H4, reference fine-tuned score) unitless · higher Uncertainty: ± 0.025 (type not stated). Table 1 does not define the spread on reference fine-tuned scores; Section 3.3 defines only the frozen-probe spread Coverage: Not reported scored / Not reported eligible	Evaluation origin not reported · Source checked Evidence concern: excluded from comparisons Table 1 prints the same frozen value and spread (.5252+-.0262) for NT on both 'Enhancer' and 'Enhancer Types', while the fine-tuned references differ (.548 and .424). One of the two frozen cells is likely a transcription error in the source; both are kept as printed and should not support comparison until the authors clarify.Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, column 'NT', rows 'Enhancer' and 'Enhancer Types' Table 1 and Table 4 disagree for HyenaDNA Splice Acceptor: Table 1 prints .4752±.0108 (recovery 50), while Table 4 prints 0.7308 ± 0.0033 (validation-selected) and 0.7360 ± 0.0052 (diagnostic grid-best). Five other HyenaDNA cells of Table 1 equal Table 4's grid-best values, but Splice Donor (.7247±.0087) equals the validation-selected value instead. Table 3 and Table 4 also name different grid-best readouts for HyenaDNA Splice All (MLP+Mean against MLP+Attn). Table 1 values are kept as printed.Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, column 'HyenaDNA', rows 'Splice Acceptor' and 'Splice Donor'; Table 4; Table 3, row 'Splice All' Methods, coverage and source NT-500M on NT benchmark H4 (reference fine-tuned) gfm2-20261011-datta2026-protocol-fine-tuned-reference Aggregation: Not reported Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, row 'H4', column 'NT', FT value
Configuration: NT-500M, fine-tuned reference score (Datta et al. 2026)	Protocol: Nucleotide Transformer benchmark, reference fine-tuned scores (Datta et al. 2026 Table 1) Dataset: Nucleotide Transformer benchmark task data (Datta et al. 2026)	0.495 MCC (NT benchmark H4ac, reference fine-tuned score) unitless · higher Uncertainty: ± 0.032 (type not stated). Table 1 does not define the spread on reference fine-tuned scores; Section 3.3 defines only the frozen-probe spread Coverage: Not reported scored / Not reported eligible	Evaluation origin not reported · Source checked Evidence concern: excluded from comparisons Table 1 prints the same frozen value and spread (.5252+-.0262) for NT on both 'Enhancer' and 'Enhancer Types', while the fine-tuned references differ (.548 and .424). One of the two frozen cells is likely a transcription error in the source; both are kept as printed and should not support comparison until the authors clarify.Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, column 'NT', rows 'Enhancer' and 'Enhancer Types' Table 1 and Table 4 disagree for HyenaDNA Splice Acceptor: Table 1 prints .4752±.0108 (recovery 50), while Table 4 prints 0.7308 ± 0.0033 (validation-selected) and 0.7360 ± 0.0052 (diagnostic grid-best). Five other HyenaDNA cells of Table 1 equal Table 4's grid-best values, but Splice Donor (.7247±.0087) equals the validation-selected value instead. Table 3 and Table 4 also name different grid-best readouts for HyenaDNA Splice All (MLP+Mean against MLP+Attn). Table 1 values are kept as printed.Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, column 'HyenaDNA', rows 'Splice Acceptor' and 'Splice Donor'; Table 4; Table 3, row 'Splice All' Methods, coverage and source NT-500M on NT benchmark H4ac (reference fine-tuned) gfm2-20261011-datta2026-protocol-fine-tuned-reference Aggregation: Not reported Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, row 'H4ac', column 'NT', FT value
Configuration: NT-500M, fine-tuned reference score (Datta et al. 2026)	Protocol: Nucleotide Transformer benchmark, reference fine-tuned scores (Datta et al. 2026 Table 1) Dataset: Nucleotide Transformer benchmark task data (Datta et al. 2026)	0.976 F1 (NT benchmark Promoter All, reference fine-tuned score) fraction · higher Uncertainty: ± 0.006 (type not stated). Table 1 does not define the spread on reference fine-tuned scores; Section 3.3 defines only the frozen-probe spread Coverage: Not reported scored / Not reported eligible	Evaluation origin not reported · Source checked Evidence concern: excluded from comparisons Table 1 prints the same frozen value and spread (.5252+-.0262) for NT on both 'Enhancer' and 'Enhancer Types', while the fine-tuned references differ (.548 and .424). One of the two frozen cells is likely a transcription error in the source; both are kept as printed and should not support comparison until the authors clarify.Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, column 'NT', rows 'Enhancer' and 'Enhancer Types' Table 1 and Table 4 disagree for HyenaDNA Splice Acceptor: Table 1 prints .4752±.0108 (recovery 50), while Table 4 prints 0.7308 ± 0.0033 (validation-selected) and 0.7360 ± 0.0052 (diagnostic grid-best). Five other HyenaDNA cells of Table 1 equal Table 4's grid-best values, but Splice Donor (.7247±.0087) equals the validation-selected value instead. Table 3 and Table 4 also name different grid-best readouts for HyenaDNA Splice All (MLP+Mean against MLP+Attn). Table 1 values are kept as printed.Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, column 'HyenaDNA', rows 'Splice Acceptor' and 'Splice Donor'; Table 4; Table 3, row 'Splice All' Methods, coverage and source NT-500M on NT benchmark Promoter All (reference fine-tuned) gfm2-20261011-datta2026-protocol-fine-tuned-reference Aggregation: Not reported Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, row 'Promoter All', column 'NT', FT value
Configuration: NT-500M, fine-tuned reference score (Datta et al. 2026)	Protocol: Nucleotide Transformer benchmark, reference fine-tuned scores (Datta et al. 2026 Table 1) Dataset: Nucleotide Transformer benchmark task data (Datta et al. 2026)	0.976 F1 (NT benchmark Promoter Non-TATA, reference fine-tuned score) fraction · higher Uncertainty: ± 0.005 (type not stated). Table 1 does not define the spread on reference fine-tuned scores; Section 3.3 defines only the frozen-probe spread Coverage: Not reported scored / Not reported eligible	Evaluation origin not reported · Source checked Evidence concern: excluded from comparisons Table 1 prints the same frozen value and spread (.5252+-.0262) for NT on both 'Enhancer' and 'Enhancer Types', while the fine-tuned references differ (.548 and .424). One of the two frozen cells is likely a transcription error in the source; both are kept as printed and should not support comparison until the authors clarify.Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, column 'NT', rows 'Enhancer' and 'Enhancer Types' Table 1 and Table 4 disagree for HyenaDNA Splice Acceptor: Table 1 prints .4752±.0108 (recovery 50), while Table 4 prints 0.7308 ± 0.0033 (validation-selected) and 0.7360 ± 0.0052 (diagnostic grid-best). Five other HyenaDNA cells of Table 1 equal Table 4's grid-best values, but Splice Donor (.7247±.0087) equals the validation-selected value instead. Table 3 and Table 4 also name different grid-best readouts for HyenaDNA Splice All (MLP+Mean against MLP+Attn). Table 1 values are kept as printed.Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, column 'HyenaDNA', rows 'Splice Acceptor' and 'Splice Donor'; Table 4; Table 3, row 'Splice All' Methods, coverage and source NT-500M on NT benchmark Promoter Non-TATA (reference fine-tuned) gfm2-20261011-datta2026-protocol-fine-tuned-reference Aggregation: Not reported Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, row 'Promoter Non-TATA', column 'NT', FT value
Configuration: NT-500M, fine-tuned reference score (Datta et al. 2026)	Protocol: Nucleotide Transformer benchmark, reference fine-tuned scores (Datta et al. 2026 Table 1) Dataset: Nucleotide Transformer benchmark task data (Datta et al. 2026)	0.966 F1 (NT benchmark Promoter TATA, reference fine-tuned score) fraction · higher Uncertainty: ± 0.013 (type not stated). Table 1 does not define the spread on reference fine-tuned scores; Section 3.3 defines only the frozen-probe spread Coverage: Not reported scored / Not reported eligible	Evaluation origin not reported · Source checked Evidence concern: excluded from comparisons Table 1 prints the same frozen value and spread (.5252+-.0262) for NT on both 'Enhancer' and 'Enhancer Types', while the fine-tuned references differ (.548 and .424). One of the two frozen cells is likely a transcription error in the source; both are kept as printed and should not support comparison until the authors clarify.Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, column 'NT', rows 'Enhancer' and 'Enhancer Types' Table 1 and Table 4 disagree for HyenaDNA Splice Acceptor: Table 1 prints .4752±.0108 (recovery 50), while Table 4 prints 0.7308 ± 0.0033 (validation-selected) and 0.7360 ± 0.0052 (diagnostic grid-best). Five other HyenaDNA cells of Table 1 equal Table 4's grid-best values, but Splice Donor (.7247±.0087) equals the validation-selected value instead. Table 3 and Table 4 also name different grid-best readouts for HyenaDNA Splice All (MLP+Mean against MLP+Attn). Table 1 values are kept as printed.Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, column 'HyenaDNA', rows 'Splice Acceptor' and 'Splice Donor'; Table 4; Table 3, row 'Splice All' Methods, coverage and source NT-500M on NT benchmark Promoter TATA (reference fine-tuned) gfm2-20261011-datta2026-protocol-fine-tuned-reference Aggregation: Not reported Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, row 'Promoter TATA', column 'NT', FT value
Configuration: NT-500M, fine-tuned reference score (Datta et al. 2026)	Protocol: Nucleotide Transformer benchmark, reference fine-tuned scores (Datta et al. 2026 Table 1) Dataset: Nucleotide Transformer benchmark task data (Datta et al. 2026)	0.981 F1 (NT benchmark Splice Acceptor, reference fine-tuned score) fraction · higher Uncertainty: ± 0.011 (type not stated). Table 1 does not define the spread on reference fine-tuned scores; Section 3.3 defines only the frozen-probe spread Coverage: Not reported scored / Not reported eligible	Evaluation origin not reported · Source checked Evidence concern: excluded from comparisons Table 1 prints the same frozen value and spread (.5252+-.0262) for NT on both 'Enhancer' and 'Enhancer Types', while the fine-tuned references differ (.548 and .424). One of the two frozen cells is likely a transcription error in the source; both are kept as printed and should not support comparison until the authors clarify.Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, column 'NT', rows 'Enhancer' and 'Enhancer Types' Table 1 and Table 4 disagree for HyenaDNA Splice Acceptor: Table 1 prints .4752±.0108 (recovery 50), while Table 4 prints 0.7308 ± 0.0033 (validation-selected) and 0.7360 ± 0.0052 (diagnostic grid-best). Five other HyenaDNA cells of Table 1 equal Table 4's grid-best values, but Splice Donor (.7247±.0087) equals the validation-selected value instead. Table 3 and Table 4 also name different grid-best readouts for HyenaDNA Splice All (MLP+Mean against MLP+Attn). Table 1 values are kept as printed.Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, column 'HyenaDNA', rows 'Splice Acceptor' and 'Splice Donor'; Table 4; Table 3, row 'Splice All' Methods, coverage and source NT-500M on NT benchmark Splice Acceptor (reference fine-tuned) gfm2-20261011-datta2026-protocol-fine-tuned-reference Aggregation: Not reported Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, row 'Splice Acceptor', column 'NT', FT value
Configuration: NT-500M, fine-tuned reference score (Datta et al. 2026)	Protocol: Nucleotide Transformer benchmark, reference fine-tuned scores (Datta et al. 2026 Table 1) Dataset: Nucleotide Transformer benchmark task data (Datta et al. 2026)	0.983 F1 (NT benchmark Splice All, reference fine-tuned score) fraction · higher Uncertainty: ± 0.008 (type not stated). Table 1 does not define the spread on reference fine-tuned scores; Section 3.3 defines only the frozen-probe spread Coverage: Not reported scored / Not reported eligible	Evaluation origin not reported · Source checked Evidence concern: excluded from comparisons Table 1 prints the same frozen value and spread (.5252+-.0262) for NT on both 'Enhancer' and 'Enhancer Types', while the fine-tuned references differ (.548 and .424). One of the two frozen cells is likely a transcription error in the source; both are kept as printed and should not support comparison until the authors clarify.Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, column 'NT', rows 'Enhancer' and 'Enhancer Types' Table 1 and Table 4 disagree for HyenaDNA Splice Acceptor: Table 1 prints .4752±.0108 (recovery 50), while Table 4 prints 0.7308 ± 0.0033 (validation-selected) and 0.7360 ± 0.0052 (diagnostic grid-best). Five other HyenaDNA cells of Table 1 equal Table 4's grid-best values, but Splice Donor (.7247±.0087) equals the validation-selected value instead. Table 3 and Table 4 also name different grid-best readouts for HyenaDNA Splice All (MLP+Mean against MLP+Attn). Table 1 values are kept as printed.Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, column 'HyenaDNA', rows 'Splice Acceptor' and 'Splice Donor'; Table 4; Table 3, row 'Splice All' Methods, coverage and source NT-500M on NT benchmark Splice All (reference fine-tuned) gfm2-20261011-datta2026-protocol-fine-tuned-reference Aggregation: Not reported Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, row 'Splice All', column 'NT', FT value
Configuration: NT-500M, fine-tuned reference score (Datta et al. 2026)	Protocol: Nucleotide Transformer benchmark, reference fine-tuned scores (Datta et al. 2026 Table 1) Dataset: Nucleotide Transformer benchmark task data (Datta et al. 2026)	0.985 F1 (NT benchmark Splice Donor, reference fine-tuned score) fraction · higher Uncertainty: ± 0.022 (type not stated). Table 1 does not define the spread on reference fine-tuned scores; Section 3.3 defines only the frozen-probe spread Coverage: Not reported scored / Not reported eligible	Evaluation origin not reported · Source checked Evidence concern: excluded from comparisons Table 1 prints the same frozen value and spread (.5252+-.0262) for NT on both 'Enhancer' and 'Enhancer Types', while the fine-tuned references differ (.548 and .424). One of the two frozen cells is likely a transcription error in the source; both are kept as printed and should not support comparison until the authors clarify.Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, column 'NT', rows 'Enhancer' and 'Enhancer Types' Table 1 and Table 4 disagree for HyenaDNA Splice Acceptor: Table 1 prints .4752±.0108 (recovery 50), while Table 4 prints 0.7308 ± 0.0033 (validation-selected) and 0.7360 ± 0.0052 (diagnostic grid-best). Five other HyenaDNA cells of Table 1 equal Table 4's grid-best values, but Splice Donor (.7247±.0087) equals the validation-selected value instead. Table 3 and Table 4 also name different grid-best readouts for HyenaDNA Splice All (MLP+Mean against MLP+Attn). Table 1 values are kept as printed.Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, column 'HyenaDNA', rows 'Splice Acceptor' and 'Splice Donor'; Table 4; Table 3, row 'Splice All' Methods, coverage and source NT-500M on NT benchmark Splice Donor (reference fine-tuned) gfm2-20261011-datta2026-protocol-fine-tuned-reference Aggregation: Not reported Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Table 1, row 'Splice Donor', column 'NT', FT value

Source checking is not independent reproduction. Release 2026-10-22-1b7cf027e4f6.

Use this configuration

How it works, versions and access

Strengths, limitations and unresolved questions

Evidence

Source checking verifies the cited claim or transcription. It does not establish independent reproduction.

Evidence table

Inspect claims, sources and review details

Trace each statement to its source and review. A context-only reference supports the record generally; it does not verify an individual field. Source checking does not reproduce an experiment.

One row per statement and cited source. Multiple citations are not independent evaluations. Shared locators are labelled explicitly.

0 evidence rows matching the loaded filters

Claims, original sources and review scope · Release 2026-10-22-1b7cf027e4f6
Property and statement	Original source and location	Review and provenance

No evidence rows match these filters. Choose another scope or clear the search.

How to read the evidence

Sources and history

Release 2026-10-22-1b7cf027e4f6 · Record review: source checked

1 source record and release history

Frozen but Not Always Accessible: A Representation Analysis of Genomic Language Models · Original source · arXiv:2608.05329v1, submitted 2026-08-05

Download this release (gzip)

Technical metadata and extraction receipts

Stable ID: gfm2-20261011-datta2026-config-nt-ft

areas: dna-genomes
contexts: research
method types: foundation_model
reported name: NT-500M
foundation model eligible: true
source locator: Section 2.2 model list; Table 1 column 'NT'
parameters: 500M parameters; fine-tuned score as reported by the benchmark or model source, protocol not printed
missing metadata: version: reason: unreported; note: Parameter count printed; no checkpoint revision or release version is given
training setting: Fine-tuned; protocol unreported in this source
provenance note: NT-500M, fine-tuned reference score (Datta et al. 2026), as evaluated in the cited comparison.

Related records

Suggest a correction