Model type
DNA state-space model; this record is the paper-specific evaluated configuration.
Configuration
Caduceus (character tokens)
This Caduceus configuration is a character-token baseline in a study of genomic tokenisation.
Sources
The impact of tokenizer selection in genomic language models · 1 Introduction (paragraph 9); 4 Discussion (paragraph 4)How it worksEvaluated procedure (conceptual)
Conceptual input–method–output guide. Check the procedure text and linked evaluation for fitted components, additional inputs and exact settings.
Sources
The impact of tokenizer selection in genomic language models · 1 Introduction (paragraph 9); Abstract (paragraph 1)At a glance
Biological inputs
DNA sequences represented by individual nucleotide characters
Sources
The impact of tokenizer selection in genomic language models · 1 Introduction (paragraph 6); 1 Introduction (paragraph 3)Outputs
Task-specific genomic classification predictions
Sources
The impact of tokenizer selection in genomic language models · 2 Materials and methods/2.6 Genomic tasks (paragraph 4); 2 Materials and methods/2.6 Genomic tasks (paragraph 6)limited source coverage · Automated source review, 2026-09-16. All specifications and missing details
Evaluations and results
Release 2026-09-17-d277315f7d76 · 1 evaluation · 1 metric row. Different protocols are not a single leaderboard.
| Metric and finding | Coverage and uncertainty | Evidence |
|---|---|---|
| Caduceus (character tokens): regulatory sequence classification Configuration: Caduceus (character tokens)Task: regulatory sequence classificationDataset: genomic benchmark categories task-category MCC across benchmark datasets Independent external evaluation · Evaluation metadata: needs review | ||
| 0.778 MCC Unit: unitless · Direction: unknown | Uncertainty: not reported in legacy extract Scored: Not reported · Eligible: Not reported | source checkedThe impact of tokenizer selection in genomic language models · Table 2, Regulatory row, Caduceus (char) MCC column Source checking is not independent reproduction. |
How it works
How the evaluated method works
A state-space genomic sequence model processes individual nucleotide tokens. The study compares character, k-mer and byte-pair schemes across downstream classifiers.
Sources
The impact of tokenizer selection in genomic language models · 1 Introduction (paragraph 9); Abstract (paragraph 1)Underlying method and version boundaries
Caduceus exposes distinct Ph and PS configurations. The documented Ph-131k checkpoint uses 16 layers, width 256 and reverse-complement data augmentation; PS implements reverse-complement equivariance. These training choices are not interchangeable.
Sources
kuleshov-group/caduceus README.md · README.md; introduction, model description, pretrained-model and usage sections at pinned revisionWhat was evaluated
The linked evaluation record identifies Caduceus (character tokens): regulatory sequence classification. Its dataset, split, adaptation and evidence origin remain attached to the reported results.
Sources
The impact of tokenizer selection in genomic language models · The named evaluation’s methods and comparison table; exact preserved evaluation IDs: evaluation-b2-genomic-tokenizer-selection-2025Strengths and limitations
Strengths and considerations
- Single-base tokens preserve resolution needed for position-sensitive tasks such as splice-site prediction.
Sources
The impact of tokenizer selection in genomic language models · 4 Discussion (paragraph 4); 4 Discussion (paragraph 2)
Limitations and conditions
- Tokenisation benefits are task-dependent; the paper reports different behaviour for SARS-CoV-2 classification and does not establish a universal best tokenizer.
Sources
The impact of tokenizer selection in genomic language models · 4 Discussion (paragraph 5); 3 Results/3.1 Tokenization differentially impacts performance of gLMs on specific benchmark tasks (paragraph 1)
Profile review details
Primary full text and the available official implementation/model documentation were inspected. Explanatory claims are source-backed; unresolved exact-configuration metadata is labelled explicitly. This is automated review, not a human review or independent benchmark reproduction.
Stable record: reported-model-d5bc536ca6f0d3Specifications
Inputs, training, access and other details
Explanatory profile: limited source coverage · Automated source review, 2026-09-16. Review applies to the cited claims; unresolved fields are listed below. Numerical results retain their own review status.
| Property | Description and evidence |
|---|---|
| Model type | DNA state-space model; this record is the paper-specific evaluated configuration.Sourceskuleshov-group/caduceus README.md · README.md model description |
| Architecture / procedure | A state-space genomic sequence model processes individual nucleotide tokens. The study compares character, k-mer and byte-pair schemes across downstream classifiers.SourcesThe impact of tokenizer selection in genomic language models · 1 Introduction (paragraph 9); Abstract (paragraph 1) |
| Biological inputs | DNA sequences represented by individual nucleotide charactersSourcesThe impact of tokenizer selection in genomic language models · 1 Introduction (paragraph 6); 1 Introduction (paragraph 3) |
| Outputs | Task-specific genomic classification predictionsSourcesThe impact of tokenizer selection in genomic language models · 2 Materials and methods/2.6 Genomic tasks (paragraph 4); 2 Materials and methods/2.6 Genomic tasks (paragraph 6) |
| Parameters | 3.9-million-parameter variant as reported for this rowSourcesThe impact of tokenizer selection in genomic language models · 2 Materials and methods/2.3 Metrics (paragraph 1); 2 Materials and methods/2.6 Genomic tasks (paragraph 6) |
| Known versions / configuration | 3.9M parameter variantSourcesThe impact of tokenizer selection in genomic language models · 2 Materials and methods/2.6 Genomic tasks (paragraph 6); 5 Conclusions (paragraph 1) |
| Training data / fitting | The tokenisation study’s model pretraining and 44 classification fine-tuning tasks; each task retains its own adaptation and split.SourcesThe impact of tokenizer selection in genomic language models · 2 Materials and methods/2.3 Metrics (paragraph 1); 4 Discussion (paragraph 5) |
| Context limits | The paper uses a 4,000-nucleotide context for the compared genomic tokenisation models.SourcesThe impact of tokenizer selection in genomic language models · 1 Introduction (paragraph 5); 1 Introduction (paragraph 9) |
| Access | Official upstream implementation and usage documentation: https://github.com/kuleshov-group/caduceus/blob/0060a6d8079b6a040fc55d505e15972a327b70a6/README.md. This pinned documentation revision is not automatically the evaluated weight revision.Sourceskuleshov-group/caduceus README.md · README.md; installation, model download and usage instructions |
| Code licence | Apache 2.0 (upstream repository code at the cited revision; this does not establish every dependency or historical checkpoint licence).Sourceskuleshov-group/caduceus LICENSE · LICENSE; complete licence text |
| Weights licence | The inspected model-access documentation does not explicitly identify terms for this exact evaluated checkpoint or fitted head; repository code terms are shown separately. · Not reported in inspected sourcesSourceskuleshov-group/caduceus README.md · README.md; checkpoint/access documentation and licence scope |
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.
20 evidence rows matching the loaded filters
| Property and statement | Original source and location | Review and provenance |
|---|---|---|
| Diagram caption Conceptual input–method–output guide. Check the procedure text and linked evaluation for fitted components, additional inputs and exact settings. Individual claims | The impact of tokenizer selection in genomic language models 1 Introduction (paragraph 9); Abstract (paragraph 1) Version: journal full text in PMC | source checked automated source review · 2026-09-16 Audit detailsPrimary full text and the available official implementation/model documentation were inspected. Explanatory claims are source-backed; unresolved exact-configuration metadata is labelled explicitly. This is automated review, not a human review or independent benchmark reproduction. Field: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
| Diagram steps ["DNA sequences represented by individual nucleotide characters","Caduceus (character tokens)","Task-specific genomic classification predictions"] Individual claims | The impact of tokenizer selection in genomic language models 1 Introduction (paragraph 9); Abstract (paragraph 1) Version: journal full text in PMC | source checked automated source review · 2026-09-16 Audit detailsPrimary full text and the available official implementation/model documentation were inspected. Explanatory claims are source-backed; unresolved exact-configuration metadata is labelled explicitly. This is automated review, not a human review or independent benchmark reproduction. Field: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
| Diagram title Evaluated procedure (conceptual) Individual claims | The impact of tokenizer selection in genomic language models 1 Introduction (paragraph 9); Abstract (paragraph 1) Version: journal full text in PMC | source checked automated source review · 2026-09-16 Audit detailsPrimary full text and the available official implementation/model documentation were inspected. Explanatory claims are source-backed; unresolved exact-configuration metadata is labelled explicitly. This is automated review, not a human review or independent benchmark reproduction. Field: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
| Model type DNA state-space model; this record is the paper-specific evaluated configuration. Individual claims | kuleshov-group/caduceus README.md README.md model description Version: 0060a6d8079b6a040fc55d505e15972a327b70a6 | source checked automated source review · 2026-09-16 Audit detailsPrimary full text and the available official implementation/model documentation were inspected. Explanatory claims are source-backed; unresolved exact-configuration metadata is labelled explicitly. This is automated review, not a human review or independent benchmark reproduction. Field: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
| Architecture / procedure A state-space genomic sequence model processes individual nucleotide tokens. The study compares character, k-mer and byte-pair schemes across downstream classifiers. Individual claims | The impact of tokenizer selection in genomic language models 1 Introduction (paragraph 9); Abstract (paragraph 1) Version: journal full text in PMC | source checked automated source review · 2026-09-16 Audit detailsPrimary full text and the available official implementation/model documentation were inspected. Explanatory claims are source-backed; unresolved exact-configuration metadata is labelled explicitly. This is automated review, not a human review or independent benchmark reproduction. Field: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
| Weights licence The inspected model-access documentation does not explicitly identify terms for this exact evaluated checkpoint or fitted head; repository code terms are shown separately. Individual claims | kuleshov-group/caduceus README.md README.md; checkpoint/access documentation and licence scope Version: 0060a6d8079b6a040fc55d505e15972a327b70a6 | unreported automated source review · 2026-09-16 Audit detailsPrimary full text and the available official implementation/model documentation were inspected. Explanatory claims are source-backed; unresolved exact-configuration metadata is labelled explicitly. This is automated review, not a human review or independent benchmark reproduction. Field: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
| Biological inputs DNA sequences represented by individual nucleotide characters Individual claims | The impact of tokenizer selection in genomic language models 1 Introduction (paragraph 6); 1 Introduction (paragraph 3) Version: journal full text in PMC | source checked automated source review · 2026-09-16 Audit detailsPrimary full text and the available official implementation/model documentation were inspected. Explanatory claims are source-backed; unresolved exact-configuration metadata is labelled explicitly. This is automated review, not a human review or independent benchmark reproduction. Field: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
| Outputs Task-specific genomic classification predictions Individual claims | The impact of tokenizer selection in genomic language models 2 Materials and methods/2.6 Genomic tasks (paragraph 4); 2 Materials and methods/2.6 Genomic tasks (paragraph 6) Version: journal full text in PMC | source checked automated source review · 2026-09-16 Audit detailsPrimary full text and the available official implementation/model documentation were inspected. Explanatory claims are source-backed; unresolved exact-configuration metadata is labelled explicitly. This is automated review, not a human review or independent benchmark reproduction. Field: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
| Parameters 3.9-million-parameter variant as reported for this row Individual claims | The impact of tokenizer selection in genomic language models 2 Materials and methods/2.3 Metrics (paragraph 1); 2 Materials and methods/2.6 Genomic tasks (paragraph 6) Version: journal full text in PMC | source checked automated source review · 2026-09-16 Audit detailsPrimary full text and the available official implementation/model documentation were inspected. Explanatory claims are source-backed; unresolved exact-configuration metadata is labelled explicitly. This is automated review, not a human review or independent benchmark reproduction. Field: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
| Known versions / configuration 3.9M parameter variant Individual claims | The impact of tokenizer selection in genomic language models 2 Materials and methods/2.6 Genomic tasks (paragraph 6); 5 Conclusions (paragraph 1) Version: journal full text in PMC | source checked automated source review · 2026-09-16 Audit detailsPrimary full text and the available official implementation/model documentation were inspected. Explanatory claims are source-backed; unresolved exact-configuration metadata is labelled explicitly. This is automated review, not a human review or independent benchmark reproduction. Field: Source artifact SHA-256: Hash scope: Hash scope not separately documented; inspect source record |
Sources and history
Release 2026-09-17-d277315f7d76 · Record review: needs review
3 source records and release history
- The impact of tokenizer selection in genomic language models · Original source · journal full text in PMC
- kuleshov-group/caduceus README.md · Original source · 0060a6d8079b6a040fc55d505e15972a327b70a6
- kuleshov-group/caduceus LICENSE · Original source · 0060a6d8079b6a040fc55d505e15972a327b70a6
Technical metadata and extraction receipts
Stable ID: reported-model-d5bc536ca6f0d3
- areas
- dna-genomes
- entity level
- method
- version
- 3.9M parameter variant
- reported name
- Caduceus (character tokens)
- historical missing metadata
- checkpoint revision: not_reported_in_legacy_extract; training data: not_reported_in_legacy_extract; licence: not_reported_in_legacy_extract
- metadata review scope
- historical_missing_metadata preserves the original discovery state. Current descriptive evidence and missingness are recorded in profile.facts; numerical-result review is separate.
- legacy kinds
- model
- entity classification
- review date: 2026-09-17; rationale: This source-scoped entry preserves the method/configuration actually named in an evaluation. It is neither a global family identity nor proof of an immutable checkpoint; the linked evaluation retains adaptation, fitting and scoring details.; source ids: genomic-tokenizer-selection-2025; evidence-reported-base-caduceus-readme-md; source locator: 1 Introduction (paragraph 9); Abstract (paragraph 1) | README.md model description | 1 Introduction (paragraph 9); 4 Discussion (paragraph 4); ambiguities: Configuration means the source-labelled evaluated identity. It does not establish missing checkpoint hashes, default settings or equivalence to same-named records in other papers.