Configuration

DNABERT-2 (fine-tuned)

DNABERT-2 learns DNA representations that can be adapted to genomic prediction tasks.

Sources (3)

MAGICS-LAB/DNABERT_2: README.md; MAGICS-LAB/DNABERT_2: LICENSE; dnabert2: Primary paper PDF · DNABERT-2 paper Sections 3.2, 4.1 and 5.2 (Further Pre-Training); official model card and implementation README; official DNABERT-2-117M checkpoint LICENSE at revision 7bce263b15377fc15361f52cfab88f8b586abda0

22 evaluations · 22 results

How it worksDNABERT-2 workflow

Conceptual summary of the documented data flow; optional inputs and configured downstream stages must be reported for a reproducible evaluation.

Sources (3)

Overview

Model type

Masked-token DNA transformer encoder

Inputs

DNA sequence tokenized with the supplied tokenizer.

Outputs

Token representations and, after a specified adaptation, task predictions.

Access

Official project documentation and implementation: https://github.com/MAGICS-LAB/DNABERT_2

Sources (3)

limited source coverage · Automated source review, 2026-09-23. All specifications and missing details

Evaluations and results

22 evaluations · 22 results. Different protocols are not a single leaderboard.

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: DNABERT-2 (fine-tuned)	Task: DART-Eval CA-AUROC-GM12878: Chromatin activity prediction, GM12878, positives against negatives Dataset subset: ENCODE chromatin accessibility peaks in five cell lines (DART-Eval split)	0.916 auroc fraction · higher Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible	Author-reported evaluation · Source checked Methods, coverage and source DNABERT-2 (fine-tuned) on DART-Eval CA-AUROC-GM12878: Chromatin activity prediction, GM12878, positives against negatives Separating positive GM12878 peaks from matched negatives. Aggregation: Not reported DART-Eval: A Comprehensive DNA Language Model Evaluation Benchmark on Regulatory DNA · Table 5, row(Fine-Tuned DNABERT-2), column(CA-AUROC-GM12878)
Configuration: DNABERT-2 (fine-tuned)	Task: DART-Eval CA-AUROC-H1ESC: Chromatin activity prediction, H1ESC, positives against negatives Dataset subset: ENCODE chromatin accessibility peaks in five cell lines (DART-Eval split)	0.94 auroc fraction · higher Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible	Author-reported evaluation · Source checked Methods, coverage and source DNABERT-2 (fine-tuned) on DART-Eval CA-AUROC-H1ESC: Chromatin activity prediction, H1ESC, positives against negatives Separating positive H1ESC peaks from matched negatives. Aggregation: Not reported DART-Eval: A Comprehensive DNA Language Model Evaluation Benchmark on Regulatory DNA · Table 5, row(Fine-Tuned DNABERT-2), column(CA-AUROC-H1ESC)
Configuration: DNABERT-2 (fine-tuned)	Task: DART-Eval CA-AUROC-HEPG2: Chromatin activity prediction, HEPG2, positives against negatives Dataset subset: ENCODE chromatin accessibility peaks in five cell lines (DART-Eval split)	0.893 auroc fraction · higher Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible	Author-reported evaluation · Source checked Methods, coverage and source DNABERT-2 (fine-tuned) on DART-Eval CA-AUROC-HEPG2: Chromatin activity prediction, HEPG2, positives against negatives Separating positive HEPG2 peaks from matched negatives. Aggregation: Not reported DART-Eval: A Comprehensive DNA Language Model Evaluation Benchmark on Regulatory DNA · Table 5, row(Fine-Tuned DNABERT-2), column(CA-AUROC-HEPG2)
Configuration: DNABERT-2 (fine-tuned)	Task: DART-Eval CA-AUROC-IMR90: Chromatin activity prediction, IMR90, positives against negatives Dataset subset: ENCODE chromatin accessibility peaks in five cell lines (DART-Eval split)	0.963 auroc fraction · higher Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible	Author-reported evaluation · Source checked Methods, coverage and source DNABERT-2 (fine-tuned) on DART-Eval CA-AUROC-IMR90: Chromatin activity prediction, IMR90, positives against negatives Separating positive IMR90 peaks from matched negatives. Aggregation: Not reported DART-Eval: A Comprehensive DNA Language Model Evaluation Benchmark on Regulatory DNA · Table 5, row(Fine-Tuned DNABERT-2), column(CA-AUROC-IMR90)
Configuration: DNABERT-2 (fine-tuned)	Task: DART-Eval CA-AUROC-K562: Chromatin activity prediction, K562, positives against negatives Dataset subset: ENCODE chromatin accessibility peaks in five cell lines (DART-Eval split)	0.917 auroc fraction · higher Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible	Author-reported evaluation · Source checked Methods, coverage and source DNABERT-2 (fine-tuned) on DART-Eval CA-AUROC-K562: Chromatin activity prediction, K562, positives against negatives Separating positive K562 peaks from matched negatives. Aggregation: Not reported DART-Eval: A Comprehensive DNA Language Model Evaluation Benchmark on Regulatory DNA · Table 5, row(Fine-Tuned DNABERT-2), column(CA-AUROC-K562)
Configuration: DNABERT-2 (fine-tuned)	Task: DART-Eval CA-SPEARMAN-GM12878: Chromatin activity prediction, GM12878, positives only Dataset subset: ENCODE chromatin accessibility peaks in five cell lines (DART-Eval split)	0.489 spearman_r correlation · higher Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible	Author-reported evaluation · Source checked Methods, coverage and source DNABERT-2 (fine-tuned) on DART-Eval CA-SPEARMAN-GM12878: Chromatin activity prediction, GM12878, positives only Rank correlation with measured accessibility among positive GM12878 peaks. Aggregation: Not reported DART-Eval: A Comprehensive DNA Language Model Evaluation Benchmark on Regulatory DNA · Table 5, row(Fine-Tuned DNABERT-2), column(CA-SPEARMAN-GM12878)
Configuration: DNABERT-2 (fine-tuned)	Task: DART-Eval CA-SPEARMAN-H1ESC: Chromatin activity prediction, H1ESC, positives only Dataset subset: ENCODE chromatin accessibility peaks in five cell lines (DART-Eval split)	0.717 spearman_r correlation · higher Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible	Author-reported evaluation · Source checked Methods, coverage and source DNABERT-2 (fine-tuned) on DART-Eval CA-SPEARMAN-H1ESC: Chromatin activity prediction, H1ESC, positives only Rank correlation with measured accessibility among positive H1ESC peaks. Aggregation: Not reported DART-Eval: A Comprehensive DNA Language Model Evaluation Benchmark on Regulatory DNA · Table 5, row(Fine-Tuned DNABERT-2), column(CA-SPEARMAN-H1ESC)
Configuration: DNABERT-2 (fine-tuned)	Task: DART-Eval CA-SPEARMAN-HEPG2: Chromatin activity prediction, HEPG2, positives only Dataset subset: ENCODE chromatin accessibility peaks in five cell lines (DART-Eval split)	0.472 spearman_r correlation · higher Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible	Author-reported evaluation · Source checked Methods, coverage and source DNABERT-2 (fine-tuned) on DART-Eval CA-SPEARMAN-HEPG2: Chromatin activity prediction, HEPG2, positives only Rank correlation with measured accessibility among positive HEPG2 peaks. Aggregation: Not reported DART-Eval: A Comprehensive DNA Language Model Evaluation Benchmark on Regulatory DNA · Table 5, row(Fine-Tuned DNABERT-2), column(CA-SPEARMAN-HEPG2)
Configuration: DNABERT-2 (fine-tuned)	Task: DART-Eval CA-SPEARMAN-IMR90: Chromatin activity prediction, IMR90, positives only Dataset subset: ENCODE chromatin accessibility peaks in five cell lines (DART-Eval split)	0.47 spearman_r correlation · higher Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible	Author-reported evaluation · Source checked Methods, coverage and source DNABERT-2 (fine-tuned) on DART-Eval CA-SPEARMAN-IMR90: Chromatin activity prediction, IMR90, positives only Rank correlation with measured accessibility among positive IMR90 peaks. Aggregation: Not reported DART-Eval: A Comprehensive DNA Language Model Evaluation Benchmark on Regulatory DNA · Table 5, row(Fine-Tuned DNABERT-2), column(CA-SPEARMAN-IMR90)
Configuration: DNABERT-2 (fine-tuned)	Task: DART-Eval CA-SPEARMAN-K562: Chromatin activity prediction, K562, positives only Dataset subset: ENCODE chromatin accessibility peaks in five cell lines (DART-Eval split)	0.529 spearman_r correlation · higher Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible	Author-reported evaluation · Source checked Methods, coverage and source DNABERT-2 (fine-tuned) on DART-Eval CA-SPEARMAN-K562: Chromatin activity prediction, K562, positives only Rank correlation with measured accessibility among positive K562 peaks. Aggregation: Not reported DART-Eval: A Comprehensive DNA Language Model Evaluation Benchmark on Regulatory DNA · Table 5, row(Fine-Tuned DNABERT-2), column(CA-SPEARMAN-K562)
Configuration: DNABERT-2 (fine-tuned)	Task: DART-Eval CTS-ACC: Cell-type-specific element classification, overall accuracy Dataset subset: ENCODE chromatin accessibility peaks in five cell lines (DART-Eval split)	0.65 accuracy fraction · higher Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible	Author-reported evaluation · Source checked Methods, coverage and source DNABERT-2 (fine-tuned) on DART-Eval CTS-ACC: Cell-type-specific element classification, overall accuracy Classify which of five cell lines a accessible element belongs to. Aggregation: Not reported DART-Eval: A Comprehensive DNA Language Model Evaluation Benchmark on Regulatory DNA · Table 4, row(Fine-Tuned DNABERT-2), column(CTS-ACC)
Configuration: DNABERT-2 (fine-tuned)	Task: DART-Eval CTS-GM12878: Cell-type-specific element classification, GM12878 Dataset subset: ENCODE chromatin accessibility peaks in five cell lines (DART-Eval split)	0.894 auroc fraction · higher Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible	Author-reported evaluation · Source checked Methods, coverage and source DNABERT-2 (fine-tuned) on DART-Eval CTS-GM12878: Cell-type-specific element classification, GM12878 One-against-rest AUROC for GM12878 accessible elements. Aggregation: Not reported DART-Eval: A Comprehensive DNA Language Model Evaluation Benchmark on Regulatory DNA · Table 4, row(Fine-Tuned DNABERT-2), column(CTS-GM12878)
Configuration: DNABERT-2 (fine-tuned)	Task: DART-Eval CTS-H1ESC: Cell-type-specific element classification, H1ESC Dataset subset: ENCODE chromatin accessibility peaks in five cell lines (DART-Eval split)	0.93 auroc fraction · higher Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible	Author-reported evaluation · Source checked Methods, coverage and source DNABERT-2 (fine-tuned) on DART-Eval CTS-H1ESC: Cell-type-specific element classification, H1ESC One-against-rest AUROC for H1ESC accessible elements. Aggregation: Not reported DART-Eval: A Comprehensive DNA Language Model Evaluation Benchmark on Regulatory DNA · Table 4, row(Fine-Tuned DNABERT-2), column(CTS-H1ESC)
Configuration: DNABERT-2 (fine-tuned)	Task: DART-Eval CTS-HEPG2: Cell-type-specific element classification, HEPG2 Dataset subset: ENCODE chromatin accessibility peaks in five cell lines (DART-Eval split)	0.891 auroc fraction · higher Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible	Author-reported evaluation · Source checked Methods, coverage and source DNABERT-2 (fine-tuned) on DART-Eval CTS-HEPG2: Cell-type-specific element classification, HEPG2 One-against-rest AUROC for HEPG2 accessible elements. Aggregation: Not reported DART-Eval: A Comprehensive DNA Language Model Evaluation Benchmark on Regulatory DNA · Table 4, row(Fine-Tuned DNABERT-2), column(CTS-HEPG2)
Configuration: DNABERT-2 (fine-tuned)	Task: DART-Eval CTS-IMR90: Cell-type-specific element classification, IMR90 Dataset subset: ENCODE chromatin accessibility peaks in five cell lines (DART-Eval split)	0.922 auroc fraction · higher Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible	Author-reported evaluation · Source checked Methods, coverage and source DNABERT-2 (fine-tuned) on DART-Eval CTS-IMR90: Cell-type-specific element classification, IMR90 One-against-rest AUROC for IMR90 accessible elements. Aggregation: Not reported DART-Eval: A Comprehensive DNA Language Model Evaluation Benchmark on Regulatory DNA · Table 4, row(Fine-Tuned DNABERT-2), column(CTS-IMR90)
Configuration: DNABERT-2 (fine-tuned)	Task: DART-Eval CTS-K562: Cell-type-specific element classification, K562 Dataset subset: ENCODE chromatin accessibility peaks in five cell lines (DART-Eval split)	0.871 auroc fraction · higher Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible	Author-reported evaluation · Source checked Methods, coverage and source DNABERT-2 (fine-tuned) on DART-Eval CTS-K562: Cell-type-specific element classification, K562 One-against-rest AUROC for K562 accessible elements. Aggregation: Not reported DART-Eval: A Comprehensive DNA Language Model Evaluation Benchmark on Regulatory DNA · Table 4, row(Fine-Tuned DNABERT-2), column(CTS-K562)
Configuration: DNABERT-2 (fine-tuned)	Task: DART-Eval REI-ABS: Regulatory element identification, absolute accuracy Dataset subset: ENCODE cCREs against dinucleotide-shuffled backgrounds (DART-Eval split)	0.913 absolute_accuracy fraction · higher Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible	Author-reported evaluation · Source checked Methods, coverage and source DNABERT-2 (fine-tuned) on DART-Eval REI-ABS: Regulatory element identification, absolute accuracy Distinguish ENCODE cCREs from dinucleotide-shuffled background sequences. The setting the model was run in is part of the method name. Aggregation: Not reported DART-Eval: A Comprehensive DNA Language Model Evaluation Benchmark on Regulatory DNA · Table 3, row(DNABERT-2), column(fine-tuned absolute_accuracy)
Configuration: DNABERT-2 (fine-tuned)	Task: DART-Eval REI-PAIR: Regulatory element identification, paired accuracy Dataset subset: ENCODE cCREs against dinucleotide-shuffled backgrounds (DART-Eval split)	0.973 paired_accuracy fraction · higher Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible	Author-reported evaluation · Source checked Methods, coverage and source DNABERT-2 (fine-tuned) on DART-Eval REI-PAIR: Regulatory element identification, paired accuracy Distinguish ENCODE cCREs from dinucleotide-shuffled background sequences. The setting the model was run in is part of the method name. Aggregation: Not reported DART-Eval: A Comprehensive DNA Language Model Evaluation Benchmark on Regulatory DNA · Table 3, row(DNABERT-2), column(fine-tuned paired_accuracy)
Configuration: DNABERT-2 (fine-tuned)	Task: DART-Eval VS-AFRICAN-AUROC: Variant scoring on Chromatin QTLs in African LCLs, AUROC Dataset subset: Chromatin QTLs in African LCLs (DART-Eval split)	0.616 auroc fraction · higher Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible	Author-reported evaluation · Source checked Methods, coverage and source DNABERT-2 (fine-tuned) on DART-Eval VS-AFRICAN-AUROC: Variant scoring on Chromatin QTLs in African LCLs, AUROC Score the effect of a variant on chromatin accessibility, against the measured QTL call. Aggregation: Not reported DART-Eval: A Comprehensive DNA Language Model Evaluation Benchmark on Regulatory DNA · Table 6, row(African DNABERT-2), column(fine-tuned auroc)
Configuration: DNABERT-2 (fine-tuned)	Task: DART-Eval VS-AFRICAN-PEARSON_R: Variant scoring on Chromatin QTLs in African LCLs, Pearson r Dataset subset: Chromatin QTLs in African LCLs (DART-Eval split)	0.184 pearson_r correlation · higher Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible	Author-reported evaluation · Source checked Methods, coverage and source DNABERT-2 (fine-tuned) on DART-Eval VS-AFRICAN-PEARSON_R: Variant scoring on Chromatin QTLs in African LCLs, Pearson r Score the effect of a variant on chromatin accessibility, against the measured QTL call. Aggregation: Not reported DART-Eval: A Comprehensive DNA Language Model Evaluation Benchmark on Regulatory DNA · Table 6, row(African DNABERT-2), column(fine-tuned pearson_r)
Configuration: DNABERT-2 (fine-tuned)	Task: DART-Eval VS-YORUBAN-AUROC: Variant scoring on DNase QTLs in Yoruban LCLs, AUROC Dataset subset: DNase QTLs in Yoruban LCLs (DART-Eval split)	0.631 auroc fraction · higher Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible	Author-reported evaluation · Source checked Methods, coverage and source DNABERT-2 (fine-tuned) on DART-Eval VS-YORUBAN-AUROC: Variant scoring on DNase QTLs in Yoruban LCLs, AUROC Score the effect of a variant on chromatin accessibility, against the measured QTL call. Aggregation: Not reported DART-Eval: A Comprehensive DNA Language Model Evaluation Benchmark on Regulatory DNA · Table 6, row(Yoruban DNABERT-2), column(fine-tuned auroc)
Configuration: DNABERT-2 (fine-tuned)	Task: DART-Eval VS-YORUBAN-PEARSON_R: Variant scoring on DNase QTLs in Yoruban LCLs, Pearson r Dataset subset: DNase QTLs in Yoruban LCLs (DART-Eval split)	0.473 pearson_r correlation · higher Uncertainty: Not reported Coverage: Not reported scored / Not reported eligible	Author-reported evaluation · Source checked Methods, coverage and source DNABERT-2 (fine-tuned) on DART-Eval VS-YORUBAN-PEARSON_R: Variant scoring on DNase QTLs in Yoruban LCLs, Pearson r Score the effect of a variant on chromatin accessibility, against the measured QTL call. Aggregation: Not reported DART-Eval: A Comprehensive DNA Language Model Evaluation Benchmark on Regulatory DNA · Table 6, row(Yoruban DNABERT-2), column(fine-tuned pearson_r)

Source checking is not independent reproduction. Release 2026-09-29-06401fd5b220.

Use this model

How it works, versions and access

Related profile: DNABERT-2. This page retains the exact record and its evaluation context.

This configuration

DNA language model evaluated by the DART-Eval authors in the fine-tuned setting.

record: DNABERT-2 (fine-tuned)
configuration: Not reported
entity type: Configuration

How it works

DNABERT-2 merges recurring DNA substrings into byte-pair tokens, then processes those tokens with a masked-language-model transformer. ALiBi supplies distance-dependent attention biases, while FlashAttention changes how attention is computed. The resulting contextual embeddings need an explicit pooling rule and prediction head for a downstream task.

Sources (3)

Versions and reproducibility

DNABERT-2-117M model card and official DNABERT_2 implementation. ALiBi permits inference beyond the pretraining sequence length, subject to attention/memory cost; this does not establish unlimited biological context or validated accuracy at arbitrary lengths.

Sources (3)

Strengths, limitations and unresolved questions

Strengths and limitations

Strengths and considerations

The released model supports embedding extraction and task-specific fine-tuning.
Sources (3)
MAGICS-LAB/DNABERT_2: README.md; MAGICS-LAB/DNABERT_2: LICENSE; dnabert2: Primary paper PDF · DNABERT-2 paper Sections 3.2, 4.1 and 5.2 (Further Pre-Training); official model card and implementation README; official DNABERT-2-117M checkpoint LICENSE at revision 7bce263b15377fc15361f52cfab88f8b586abda0

Limitations and conditions

An embedding model alone is not the same evaluated pipeline as frozen embeddings followed by logistic regression. Tokenization and pooling choices must be preserved.
Sources (3)
MAGICS-LAB/DNABERT_2: README.md; MAGICS-LAB/DNABERT_2: LICENSE; dnabert2: Primary paper PDF · DNABERT-2 paper Sections 3.2, 4.1 and 5.2 (Further Pre-Training); official model card and implementation README; official DNABERT-2-117M checkpoint LICENSE at revision 7bce263b15377fc15361f52cfab88f8b586abda0

Profile review details

Follow-up review of Reference checkpoint, Context limits, Training cutoff, Further pretraining. Source locations and before/after decisions are recorded in the 23 September profile-evidence audit. Other explanatory content retains its earlier source scope. No human scientific review or independent reproduction is implied.

Stable record: discovery-model-dnabert-2

Specifications

Inputs, training, access and other details

Explanatory profile: limited source coverage · Automated source review, 2026-09-23. Review applies to the cited claims; unresolved fields are listed below. Numerical results retain their own review status.

Inputs, outputs and configuration
Property	Description and evidence
Model type	Masked-token DNA transformer encoder Sources (3) MAGICS-LAB/DNABERT_2: README.md; MAGICS-LAB/DNABERT_2: LICENSE; dnabert2: Primary paper PDF · DNABERT-2 paper Sections 3.2, 4.1 and 5.2 (Further Pre-Training); official model card and implementation README; official DNABERT-2-117M checkpoint LICENSE at revision 7bce263b15377fc15361f52cfab88f8b586abda0
Architecture	BERT-style DNA encoder with byte-pair tokenization, ALiBi relative attention biases and FlashAttention; task heads and pooling are separately configured. Sources (3) MAGICS-LAB/DNABERT_2: README.md; MAGICS-LAB/DNABERT_2: LICENSE; dnabert2: Primary paper PDF · DNABERT-2 paper Sections 3.2, 4.1 and 5.2 (Further Pre-Training); official model card and implementation README; official DNABERT-2-117M checkpoint LICENSE at revision 7bce263b15377fc15361f52cfab88f8b586abda0
Inputs	DNA sequence tokenized with the supplied tokenizer. Sources (3) MAGICS-LAB/DNABERT_2: README.md; MAGICS-LAB/DNABERT_2: LICENSE; dnabert2: Primary paper PDF · DNABERT-2 paper Sections 3.2, 4.1 and 5.2 (Further Pre-Training); official model card and implementation README; official DNABERT-2-117M checkpoint LICENSE at revision 7bce263b15377fc15361f52cfab88f8b586abda0
Outputs	Token representations and, after a specified adaptation, task predictions. Sources (3) MAGICS-LAB/DNABERT_2: README.md; MAGICS-LAB/DNABERT_2: LICENSE; dnabert2: Primary paper PDF · DNABERT-2 paper Sections 3.2, 4.1 and 5.2 (Further Pre-Training); official model card and implementation README; official DNABERT-2-117M checkpoint LICENSE at revision 7bce263b15377fc15361f52cfab88f8b586abda0
Parameters	117 million for DNABERT-2-117M; family names do not establish a particular checkpoint. Sources (3) MAGICS-LAB/DNABERT_2: README.md; MAGICS-LAB/DNABERT_2: LICENSE; dnabert2: Primary paper PDF · DNABERT-2 paper Sections 3.2, 4.1 and 5.2 (Further Pre-Training); official model card and implementation README; official DNABERT-2-117M checkpoint LICENSE at revision 7bce263b15377fc15361f52cfab88f8b586abda0
Known versions	DNABERT-2-117M model card and official DNABERT_2 implementation. Sources (3) MAGICS-LAB/DNABERT_2: README.md; MAGICS-LAB/DNABERT_2: LICENSE; dnabert2: Primary paper PDF · DNABERT-2 paper Sections 3.2, 4.1 and 5.2 (Further Pre-Training); official model card and implementation README; official DNABERT-2-117M checkpoint LICENSE at revision 7bce263b15377fc15361f52cfab88f8b586abda0
Training data	The paper describes a 32.49-billion-base corpus covering 135 species in six groups, alongside a 2.75-billion-base human corpus. Further GUE-domain pretraining is a separately reported model variant. Sources (3) MAGICS-LAB/DNABERT_2: README.md; MAGICS-LAB/DNABERT_2: LICENSE; dnabert2: Primary paper PDF · DNABERT-2 paper Sections 3.2, 4.1 and 5.2 (Further Pre-Training); official model card and implementation README; official DNABERT-2-117M checkpoint LICENSE at revision 7bce263b15377fc15361f52cfab88f8b586abda0
Training cutoff	Section4.1 identifies the human and 135-species corpora but does not establish a latest-sequence deposition date. The model repository revision is not a training-data cutoff. · Not reported in inspected sources Sources dnabert2 paper v2: primary artifact · Section4.1; Table11
Context limits	The paper describes pretraining on 700-base sequences and evaluates 5,000–10,000-base GUE+ inputs after fine-tuning. This does not establish frozen-model accuracy at those lengths; tokenisation, truncation and adaptation must be specified. Sources dnabert2 paper v2: primary artifact · Section5.4 Results on GUE+, printed p10; Table2 input lengths
Weights licence	The official zhihan1996/DNABERT-2-117M checkpoint repository carries Apache-2.0 in its pinned LICENSE. This does not assign terms to a separately fitted downstream predictor. Sources (3) MAGICS-LAB/DNABERT_2: README.md; MAGICS-LAB/DNABERT_2: LICENSE; dnabert2: Primary paper PDF · DNABERT-2 paper Sections 3.2, 4.1 and 5.2 (Further Pre-Training); official model card and implementation README; official DNABERT-2-117M checkpoint LICENSE at revision 7bce263b15377fc15361f52cfab88f8b586abda0
Access	Official project documentation and implementation: https://github.com/MAGICS-LAB/DNABERT_2 Sources (3) MAGICS-LAB/DNABERT_2: README.md; MAGICS-LAB/DNABERT_2: LICENSE; dnabert2: Primary paper PDF · DNABERT-2 paper Sections 3.2, 4.1 and 5.2 (Further Pre-Training); official model card and implementation README; official DNABERT-2-117M checkpoint LICENSE at revision 7bce263b15377fc15361f52cfab88f8b586abda0
Code licence	Apache-2.0 Sources MAGICS-LAB/DNABERT_2: LICENSE · LICENSE: licence text
Reference checkpoint	DNABERT-2-117M: Hugging Face revision 7bce263b15377fc15361f52cfab88f8b586abda0. Its pytorch_model.bin has registry-reported SHA-256 7ff39ec77a484dd01070a41bfd6e95cdd7247bec80fe357ab43a4be33687aeba. The weight file was not downloaded for this review. Sources dnabert2 release: primary artifact · sha; siblings[pytorch_model.bin].lfs.sha256
Further pretraining	The diamond-marked DNABERT-2 variant receives additional masked-language-model training on GUE training sets. It must be distinguished from the base pretrained model in comparisons. Sources dnabert2 paper v2: primary artifact · Section5.2FurtherPreTraining; Table3caption

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.

2 evidence rows matching the loaded filters

Claims, original sources and review scope · Release 2026-09-29-06401fd5b220
Property and statement	Original source and location	Review and provenance
Relationship: family discovery-model-dnabert-2 Individual claims	DART-Eval: A Comprehensive DNA Language Model Evaluation Benchmark on Regulatory DNA Original source ↗ DART-Eval Table 1 (DNABERT-2, 117M) and adaptation-specific result tables; source-labelled configuration DNABERT-2 (fine-tuned) \| Existing reviewed locator: Table 1, row(DNABERT-2) Shared locator for this statement’s cited sources; not a separate locator for each citation. Version: 2412.05430v1 Retrieved: 2026-09-17T07:56:09.182117+00:00	source checked automated source review · 2026-09-23 Audit details Source review establishes this relationship only. Exact evaluated configurations and original numerical review status remain unchanged. Official paper fixes backbone identity and separates fine-tuning/probing/zero-shot conditions. Family link does not pool scores. Field: `links:family:discovery-model-dnabert-2` Claim: model-evaluation-identity-f7085835c81baf43724c Source artifact SHA-256: `4194b137ba55c9a2c269d119a9afec6ae1bb0feaf17d91433ae483c41221a56b` Hash scope: Exact retrieved primary paper artifact bytes. Inspected artifact
Relationship: family discovery-model-dnabert-2 Individual claims	MAGICS-LAB/DNABERT_2: README.md Original source ↗ DART-Eval Table 1 (DNABERT-2, 117M) and adaptation-specific result tables; source-labelled configuration DNABERT-2 (fine-tuned) \| Existing reviewed locator: Table 1, row(DNABERT-2) Shared locator for this statement’s cited sources; not a separate locator for each citation. Version: f25bed9ee20db966dff39e5c1571249d04e36404 Retrieved: 2026-09-16T19:46:17.892989+00:00	source checked automated source review · 2026-09-23 Audit details Source review establishes this relationship only. Exact evaluated configurations and original numerical review status remain unchanged. Official paper fixes backbone identity and separates fine-tuning/probing/zero-shot conditions. Family link does not pool scores. Field: `links:family:discovery-model-dnabert-2` Claim: model-evaluation-identity-f7085835c81baf43724c Source artifact SHA-256: `734a8cec5f667d74d421bf3b273ad7e256216109636da45aa7ceba21cd34de16` Hash scope: SHA-256 of retrieved original artifact bytes Format: original_artifact Inspected artifact

All evidence (CSV)All evidence (JSONL)How to read the evidence

Sources and history

View linked audit checks and correction history

Release 2026-09-29-06401fd5b220 · Record review: source checked

1 source records and release history

DART-Eval: A Comprehensive DNA Language Model Evaluation Benchmark on Regulatory DNA · Original source · 2412.05430v1

Download this release

Technical metadata and extraction receipts

Stable ID: dart-eval-method-dnabert-2-fine-tuned

areas: dna-genomes
source locator: Table 1, row(DNABERT-2)
missing metadata: checkpoint revision: unreported; parameters: unextracted

Related records

Suggest a correction