living-models/Botanic1-pretraining

Botanic1 pretraining data

This dataset accompanies the Botanic1 technical report. It preserves the genomic sequence shards used to pretrain the Botanic1 models, including the original sequence case, augmentation margins, train/test separation, source manifests, and assembly metadata.

Release contents

The initial release contains the main 8 kbp corpus used by Botanic1-S, M, L, and XL. The context extension corpora at 16, 32, 64, and 128 kbp are planned for this same repository and are not included in the initial upload.

Configuration	Nominal window (bp)	Training windows	Effective training base pairs
8k	8,192	7,207,506	59,043,889,152

The 8k source manifest records 322 training and four held-out genomes. Of the training species, 320 contribute at least one window. Effective base pairs are nominal window length times the manifest's training window count; they exclude augmentation margins and differ from the number of tokens processed during model training.

The 8k training corpus was shuffled before being split into its eight shards, with a buffer that covered all training windows.

Format and loading

The original files are Zstandard-compressed JSON Lines. Each line stores a text sequence and a species_id containing the species name. Lowercase letters preserve the repeat soft mask from the original sequencing data used for loss weighting. Do not uppercase the sequence before recovering this mask. Stored 8k windows have an additional 100 bp margin used for random crop augmentation. The model's 8,192 token input includes a class token and 8,191 nucleotide tokens.

Download the original shards with the Hugging Face client:

from huggingface_hub import snapshot_download

root = snapshot_download(
    repo_id="living-models/Botanic1-pretraining",
    repo_type="dataset",
    allow_patterns=["8k/*"],
    token=True,
)

Read one shard without decompressing it to disk:

import io
import json
from pathlib import Path
import zstandard

path = Path(root) / "8k/train_chunk0.jsonl.zstd"
with path.open("rb") as compressed:
    with zstandard.ZstdDecompressor().stream_reader(compressed) as stream:
        with io.TextIOWrapper(stream, encoding="utf-8") as lines:
            example = json.loads(next(lines))
            sequence = example["text"]

Loading with multiple workers

The example reads the stored order within each shard directly. Each new pass reuses that stored shuffle.

After downloading all eight training shards, use Hugging Face Datasets to stream the compressed JSONL files into a PyTorch DataLoader. Datasets assigns distinct shards to each worker, so workers do not repeat one another's records. Four workers read two shards each in this example. See the Datasets streaming guide.

This example uses datasets==5.0.1, torch, and zstandard. Save this as load_pretraining.py and run python load_pretraining.py /path/to/downloaded/snapshot using the snapshot directory returned above:

import sys
from pathlib import Path

from datasets import load_dataset
from torch.utils.data import DataLoader


def main(root):
    files = [
        str(Path(root) / f"8k/train_chunk{i}.jsonl.zstd")
        for i in range(8)
    ]
    train = load_dataset(
        "json", data_files={"train": files}, split="train", streaming=True,
    )
    assert train.num_shards == 8
    loader = DataLoader(
        train, batch_size=8, num_workers=4,
        drop_last=False, persistent_workers=False,
    )
    for epoch in range(2):
        for batch in loader:
            sequences = batch["text"]        # List of strings; case preserved.
            species = batch["species_id"]
            # Recover repeat masks, crop, tokenize, and apply MLM masking here.


if __name__ == "__main__":
    main(sys.argv[1])

Keep the if __name__ == "__main__" entry-point guard when using worker processes, including on macOS and Windows. With eight compressed training shards, up to eight workers can read distinct files in a single training process. For more details, see PyTorch data loading and Datasets with PyTorch for distributed training.

Layout and provenance

Each configuration contains eight training shards and eight test shards. release_manifest.json within each configuration lists the included files and their SHA-256 checksums.

The configuration's provenance/ directory contains the original sampling manifest, assembly catalogue, and annotation quality audit. Sequence records include species names but no genomic coordinates; provenance is available at the assembly level.

Sources and release terms

The corpus derives from plant reference assemblies obtained through NCBI and Ensembl, as documented in the technical report's source catalogue. The report describes the assembly filters, functional-region sampling, genome-length weighting, MinHash filtering, and species holdout procedure.

This dataset is released under the Living Models Research License, the same license as the Botanic1 models.

Citation

If you use this dataset in your research, please cite the technical report:

@article{Barozet2026.09.04.749355,
    author = {Barozet, Am{\'e}lie and Cabeli, Vincent and Ogier du Terrail, Jean
              and Rukhovich, Alexey and Janssoone, Thomas and Klajer, Gary
              and Sheikhitarghi, Zeinab and Andrews, Gregory and Veran, Cyril
              and Strouk, L{\'e}onard},
    title = {BOTANIC-1: a series of long-context plant genomic foundation
             models in the agentic era},
    journal = {bioRxiv},
    year = {2026},
    elocation-id = {2026.09.04.749355},
    doi = {10.64898/2026.09.04.749355},
    publisher = {Cold Spring Harbor Laboratory},
    URL = {https://www.biorxiv.org/content/early/2026/09/09/2026.09.04.749355},
    eprint = {https://www.biorxiv.org/content/early/2026/09/09/2026.09.04.749355.full.pdf},
}