# GENEB: A Benchmark for Choosing Genomic Foundation Models [![ICML 2026](https://img.shields.io/badge/ICML-2026-1f6feb.svg)](https://arxiv.org/abs/2606.04525) [![arXiv](https://img.shields.io/badge/arXiv-2606.04525-b31b1b.svg)](https://arxiv.org/abs/2606.04525) [![Leaderboard](https://img.shields.io/badge/%F0%9F%A4%97-Leaderboard-yellow.svg)](https://huggingface.co/spaces/darlednik/geneb-leaderboard) [![Dataset](https://img.shields.io/badge/%F0%9F%A4%97-Task%20data-yellow.svg)](https://huggingface.co/datasets/darlednik/geneb-tasks) [![License](https://img.shields.io/badge/license-Apache%202.0-green.svg)](LICENSE) **GENEB compares genomic foundation models and DNA language models on downstream genomics tasks. The leaderboard is organized by biological task, data regime, and model size, so it can be used to choose a model for a specific setting.** GENEB covers **40 models**, **100 DNA classification tasks**, and **13 functional categories**, including promoters, enhancers, histone modifications, DNA methylation, splice sites, and lncRNA. Every model is evaluated with the same frozen-embedding protocol in full-data, 10-shot, and 1-shot settings. > Model rankings change across biological tasks and data regimes. The overall leaderboard does not identify the best model for every use case. ๐Ÿ† [**Open the interactive leaderboard**](https://huggingface.co/spaces/darlednik/geneb-leaderboard) ยท ๐Ÿ“„ [**Read the ICML 2026 paper**](https://arxiv.org/abs/2606.04525) ยท ๐Ÿงฌ [**Download the benchmark tasks**](https://huggingface.co/datasets/darlednik/geneb-tasks) ยท โž• [**Evaluate and submit your model**](CONTRIBUTING.md) image > **Reproducing a specific model?** Reference extractors for the leaderboard models live on > the [`dev`](https://github.com/darlednik/GENEB/tree/dev) branch. Some of them need > model-specific dependencies, so `main` deliberately keeps only a minimal reference set > under `harness/extractors/` and `dev` holds the full collection. --- ## Using GENEB ### Choose a model for a downstream task Use the [interactive leaderboard](https://huggingface.co/spaces/darlednik/geneb-leaderboard) to compare models by: - biological category or individual task - full-data, 10-shot, or 1-shot performance - MCC, accuracy, or macro-F1 - parameter count and compute budget A model that ranks highly overall may perform poorly on a particular category. The leaderboard therefore reports results at the overall, category, and task levels. ### Compare genomic foundation models under one protocol Published results often use different datasets, splits, metrics, and fine-tuning procedures. GENEB evaluates all models on the same task suite with the same linear probe over frozen representations. ### Evaluate a new model The evaluation harness runs locally, so model weights and embeddings do not leave your machine. A submission contains the extractor code, metrics, and a model card. --- ## Benchmark | | | |---|---| | Genomic foundation models | 40 | | DNA classification tasks | 100 | | Functional categories | 13 | | Data regimes | full, 10-shot, 1-shot | | Primary metric | Matthews correlation coefficient (MCC) | | Evaluation | logistic regression on frozen sequence embeddings | | Repetitions | 5 random seeds | | Outputs | overall, category-level, and task-level rankings | Use the category-level and task-level tables to choose a model for a specific genomics task. --- ## Key findings **1. Aggregate leaderboards are unstable.** Model rankings vary sharply across task categories, so a single headline number hides which model you should actually use. Within the top 6 by macro MCC, GenomeOcean-4B ranks 4th overall yet is the strongest of that group on histone modification (0.545), and NT-2.5B-MS ranks 5th yet is the strongest on splice sites (0.652). **2. Scale gives only modest and inconsistent gains.** Parameter count is a weak predictor of representation quality: neither 7B model in the suite reaches the top 6, and the largest model that does โ€” GenomeOcean-4B, macro 0.573 โ€” scores below both the 3B and the 1.2B GENERator-Eukaryote models (0.605 / 0.581). **3. Architecture and pretraining alignment often outweigh parameter count.** Under matched comparison across architecture, tokenization, and pretraining data, these factors frequently matter more for representation quality than model size. **4. Category difficulty is highly uneven.** Across those same six models, promoter tasks reach 0.739โ€“0.774 MCC while DNA methylation stays at 0.231โ€“0.440 โ€” collapsing the two into one score discards the information a practitioner needs. *(Full-shot MCC. See the [paper](https://arxiv.org/abs/2606.04525) for per-category scaling correlations, non-linear probe checks, and few-shot sensitivity analyses.)* --- ## Leaderboard โ€” top 6 of 40 Full-shot, MCC. **Macro** = mean of the 13 category scores (each weighs 1/13); **Micro** = mean over all 100 tasks. | # | Model | Params | Macro | Micro | |---|-------|--------|-------|-------| | 1 | GENERator-eukaryote-3b | 3B | **0.605** | 0.598 | | 2 | LucaOne-default-step36M | 2B | 0.589 | 0.579 | | 3 | GENERator-eukaryote-1.2b | 1.2B | 0.581 | 0.581 | | 4 | GenomeOcean-4B | 4B | 0.573 | 0.563 | | 5 | NT-2.5B-MS | 2.5B | 0.573 | 0.561 | | 6 | Omni-DNA-1B | 1B | 0.568 | 0.560 | The same six models across 6 of the 13 categories โ€” note how the ordering changes column to column. Numbers in headers are task counts; **bold** marks the best of these six models in a column, not the best of all 40. | Model | Histone mod. (30) | Promoters (22) | Enhancers (8) | DNA methyl. (8) | Splice sites (7) | lncRNA (6) | |-------|------|------|------|------|------|------| | GENERator-eukaryote-3b | 0.537 | **0.774** | **0.488** | **0.440** | 0.648 | 0.453 | | LucaOne-default-step36M | 0.533 | 0.746 | 0.479 | 0.323 | 0.636 | **0.508** | | GENERator-eukaryote-1.2b | 0.521 | 0.768 | 0.485 | 0.397 | 0.629 | 0.438 | | GenomeOcean-4B | **0.545** | 0.739 | **0.488** | 0.231 | 0.580 | 0.396 | | NT-2.5B-MS | 0.498 | 0.739 | 0.487 | 0.356 | **0.652** | 0.372 | | Omni-DNA-1B | 0.505 | 0.759 | 0.474 | 0.290 | 0.604 | 0.323 | **[โ†’ All 40 models, all 13 categories, all 100 tasks, in 1-/10-/full-shot](https://huggingface.co/spaces/darlednik/geneb-leaderboard)** โ€” filter by regime, metric (MCC / accuracy / macro-F1), and model size. --- ## Evaluate your model Evaluation runs on your hardware. Model weights and intermediate embeddings stay local; a submission contributes only the extractor code, the final metrics, and a model card. ```bash # 1. Get the pinned task data revision python3 -m pip install "huggingface_hub>=0.24" python3 tools/sync_geneb_dataset.py download --local_dir ./GENEB_data # 2. Run the harness (start with --limit 5 for a smoke test) python3 harness/run_GENEB.py \ --extractor MyModelExtractor --module my_model \ --name_model org/my-model \ --model_id my-model-300m --display "My Model 300M" --params 300000000 \ --data_dir ./GENEB_data --device cuda --limit 5 # 3. Validate the produced submission file python3 tools/validate_submission.py submissions/my-model-300m.json ``` An extractor is a small wrapper that loads your encoder, prepares DNA sequences, computes hidden states, applies a pooling strategy, and returns one embedding vector per sequence. See [CONTRIBUTING.md](CONTRIBUTING.md) for evaluation and submission instructions. --- ## Evaluation protocol | | | |---|---| | Models evaluated | 40 genomic foundation models | | Tasks | 100 DNA classification tasks in 13 functional categories | | Data format | one CSV per task with columns `text`, `label`, `split` (`train` / `test`) | | Evaluation | logistic-regression probe on frozen embeddings (`max_iter=1000`) | | Regimes | `full`, `10shot`, `1shot` | | Metrics | MCC (primary; robust to class imbalance), accuracy, macro-F1 | | Seeds | `13, 17, 42, 123, 997` โ€” results averaged over all five | | Subsampling | tasks exceeding 10โต sequences are subsampled; MCC was shown empirically to stabilize beyond this size | | Data pin | dataset revision fixed in `benchmark/benchmark_spec.json` | Frozen representations isolate representation quality and keep comparisons controlled across architectures and training regimes. Released train/test partitions are used as published, without relabeling or resplitting; encoders are not fine-tuned on task labels beyond the defined linear probe. The authors declare no financial conflicts of interest: none of the 40 evaluated models was developed by the authors or their funders. --- ## Models covered
All 40 evaluated models (architecture ยท tokenization ยท params ยท pretraining data) | Model | Architecture | Tokenization | Params | Pretraining data | |-------|--------------|--------------|--------|------------------| | METAGENE-1 | T-dec | BPE | 7B | multi-species | | Evo-1-131k | StripedHyena | SN | 7B | prokaryotic | | GenomeOcean-4B | T-dec | BPE | 4B | multi-species | | GENERator-Eukaryote-3B | T-dec | k-mer | 3B | eukaryotic-genes | | DNA-GPT-3B-M | T-dec | k-mer | 3B | multi-species | | NT-2.5B-MS | T-enc | k-mer | 2.5B | multi-species | | LucaOne | T-enc | mixed | 2B | multi-species | | GENERator-Eukaryote-1.2B | T-dec | k-mer | 1.2B | eukaryotic-genes | | Omni-DNA-1B | T-dec | BPE | 1B | multi-species | | Agro-NT-1B | T-enc | k-mer | 1B | plant-genomes | | SPACE | CNN-Transformer-MoE | SN | 589M | human-mouse-profiles | | eccDNAMamba | Mamba | BPE | 537M | multi-species | | GenomeOcean-500M | T-dec | BPE | 500M | multi-species | | GENA-LM-Large-T2T | T-enc | BPE | 336M | multi-species | | Omni-DNA-300M | T-dec | BPE | 300M | multi-species | | BioFM-265M | T-dec | BioToken | 265M | human | | Enformer | CNN-Transformer | SN | 252M | human-mouse-profiles | | NT-v2-250M-MS | T-enc | k-mer | 250M | multi-species | | PlantCaduceus | Mamba | SN | 225M | plant-genomes | | OmniNA-220M | T-dec | BPE | 220M | multi-species | | GPT2-Gene-Multi-v2 | T-dec | BPE | 200M | human | | GPT2-Gene-v1 | T-dec | BPE | 200M | human | | Genomics-FM | T-enc | BPE+k-mer | 120M | multi-species | | DNABERT-S | T-enc | k-mer | 117M | multi-species-microbial | | DNABERT-2 | T-enc | BPE | 117M | multi-species | | GENA-LM-T2T-Multi | T-enc | BPE | 110M | multi-species | | GENA-LM | T-enc | BPE | 110M | human | | DNA-GPT-0.1B-H | T-dec | k-mer | 100M | multi-species | | NT-v2-100M-MS | T-enc | k-mer | 100M | multi-species | | GROVER | T-enc | BPE | 87M | human | | MutBERT | T-enc | SN | 86M | human | | DeepGene | T-enc+Graph-Transformer | BPE | 85M | human | | HyenaDNA-Large-1M | Hyena | SN | 55M | human | | NT-v2-50M-3mer-MS | T-enc | k-mer | 50M | multi-species | | NT-v2-50M-MS | T-enc | k-mer | 50M | multi-species | | HyenaDNA-Medium-160k | Hyena | SN | 14M | human | | Caduceus-PS-131k | Mamba | SN | 8M | human | | JanusDNA-72-w | Hybrid-Mamba-MoE | SN | 2M | human | | JanusDNA-72-wo | Hybrid-Mamba-MoE | SN | 2M | human | | Caduceus-PH-1k | Mamba | SN | 2M | human |
--- ## Repository overview Three coordinated parts: 1. **Evaluation harness** (`harness/`) โ€” computes embeddings, trains the protocol-defined logistic-regression probe, and emits a submission file. 2. **Benchmark definition** (`benchmark/`) โ€” task list, category map, probe settings, and model registry metadata. 3. **Leaderboard artifacts** (`leaderboard/`) โ€” static site and JSON tables built by CI from reviewed submissions; not edited manually. Baseline and contributed results live in `submissions/.json`. The repository stores extractor code and metrics, but not third-party model weights.
Repository layout ```text benchmark/ benchmark_spec.json # task list, categories, metrics, probe protocol, dataset pin model_meta.json # display names, parameter counts, links, provenance labels submissions/ .json # per-model results (reviewed before merge) leaderboard/ index.html # Hugging Face Space UI leaderboard.json # macro scores by functional category (generated) leaderboard_tasks.json # per-task scores (generated) README.md # Space metadata tools/ validate_submission.py # schema and completeness checks (CI) build_leaderboard.py # aggregate submissions into leaderboard JSON pack_submission.py # convert legacy per-task JSON logs to submission format sync_geneb_dataset.py # download or publish pinned task CSVs on Hugging Face model_cards/ .md # optional training-data and disclosure notes harness/ run_GENEB.py # end-to-end local evaluation โ†’ submission JSON extractors/ base.py # extractor interface .py # model-specific embedding module (per submission) .github/workflows/ validate.yml # PR checks on submissions and spec build-and-sync.yml # rebuild leaderboard and push to the Space (on merge to main) ```
--- ## From model to leaderboard row A contributor implements an extractor under `harness/extractors/`, runs `harness/run_GENEB.py` on the pinned task data, and opens a pull request with the submission file, the extractor module, and a model card. CI validates schema and completeness; maintainers review the extractor and metadata for plausibility; on merge, the leaderboard is rebuilt and synced to the Hugging Face Space automatically. --- ## Evaluation and reproducibility GENEB does **not** provide a central re-scoring service. Contributors run evaluation locally using the published harness and the pinned dataset revision. Maintainers review pull requests for schema compliance, completeness, and plausibility, but do not re-run every model on maintainer infrastructure. Community submissions are marked `self-reported` in metadata. | Mechanism | Role | |-----------|------| | Dataset revision in `benchmark_spec.json` | Fixes the exact benchmark data used for evaluation | | Fixed probe protocol | Keeps downstream evaluation comparable across models | | Submission schema + CI validation | Checks completeness, metric ranges, and file consistency | | Extractor code in the PR | Shows how embeddings were produced and helps others reproduce the run | | Model card | Documents architecture, training data, and possible benchmark overlap | | Provenance metadata | Marks externally submitted runs as self-reported | A run is considered reproducible when the submitted extractor, the GENEB harness version, and the pinned dataset revision are sufficient for another user to repeat the evaluation. --- ## Citation Accepted at the **43rd International Conference on Machine Learning (ICML 2026)**, Seoul, South Korea. Please cite the conference version: ```bibtex @inproceedings{ledneva2026geneb, title = {{GENEB}: Why Genomic Models Are Hard to Compare}, author = {Ledneva, Daria and Nuridinov, Mikhail and Kuznetsov, Denis}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, series = {Proceedings of Machine Learning Research}, volume = {306}, year = {2026}, publisher = {PMLR}, note = {To appear} } ```
arXiv preprint entry ```bibtex @misc{ledneva2026geneb-arxiv, title = {{GENEB}: Why Genomic Models Are Hard to Compare}, author = {Ledneva, Daria and Nuridinov, Mikhail and Kuznetsov, Denis}, year = {2026}, eprint = {2606.04525}, archivePrefix = {arXiv}, primaryClass = {cs.CL}, url = {https://arxiv.org/abs/2606.04525} } ```
--- ## License Code in this repository is released under the [Apache License 2.0](LICENSE). The license covers the harness, tools, benchmark definition, and leaderboard code. It does **not** cover third-party model weights, and it does not override the licensing terms of the upstream sources from which GENEB tasks are derived. --- ## Links - **Paper:** https://arxiv.org/abs/2606.04525 - **Leaderboard:** https://huggingface.co/spaces/darlednik/geneb-leaderboard - **Task data:** https://huggingface.co/datasets/darlednik/geneb-tasks - **Submit a model:** [CONTRIBUTING.md](CONTRIBUTING.md) Reference extractors for additional leaderboard models are currently on the [`dev`](https://github.com/darlednik/geneb/tree/dev) branch and are being integrated into `main`; `main` currently ships a minimal reference set under `harness/extractors/`. ## Contact Repository and leaderboard contact: [Daria Ledneva](mailto:a.ledn2026@gmail.com). Questions, suggestions, feedback, and model submissions are welcome.