OpenGenome2-small (og2_small)

An 11.8B-token, per-base-uniform sample of OpenGenome2, the ~8.8T-token DNA corpus used to train Evo 2, with the same mix of eukaryotic genomes, prokaryotic genomes, metagenomes, transcripts, viruses and regulatory sequence, plus a validation set of species never seen in training.

Train	11.78B bases in 1.54M windows (≤ 16,384 bp), ~39.3k species, 226 phyla
Held-out species	124M bases, 367 species absent from train (1% of species)
Evo 2 valid	99M bases from OpenGenome2's own validation split
Baseline	1.940 bits/base on held-out species (order-7 Markov model)
Source	arcinstitute/opengenome2, revision 84d2a7e, sampling rate 0.131%
License	Apache 2.0 (as OpenGenome2)

Quick start

hf download AutomatedScientist/og2_small --repo-type dataset --local-dir og2_small
cd og2_small && pip install -r scripts/requirements.txt torch

import sys; sys.path.insert(0, "scripts")
import torch.nn.functional as F
from og2dataloader import RandomCrops, EvalTiles, make_loader

train = make_loader(RandomCrops(".", "train", ctx=4096, seed=0, rc_prob=0.5), batch_size=16, num_workers=4)
heldout = make_loader(EvalTiles(".", "heldout", ctx=4096), batch_size=16)

batch = next(iter(train))  # input_ids, labels, loss_mask, lowercase: each [16, 4096]
loss = F.cross_entropy(logits.flatten(0, 1), batch["labels"].flatten(), ignore_index=-100)

The data are numpy arrays with JSONL indexes (one byte per base), read with the scripts in scripts/; there is no Parquet copy, so datasets.load_dataset and the dataset viewer do not apply.

Splits

split	bases	windows	species	what it measures
train	11.78B	1,539,679	~39.3k	training
heldout	124M	15,940	367	generalization to unseen species
valid	99M	7,811	59	Evo 2's validation split (comparability)

• heldout is drawn from the same source shards as train: every record of ~1% of species (chosen by a hash of the species name) and ~1% of untagged records, sampled at the same rate. It shares no species with train and has the same subset mix, so it is the split to report.
• valid samples OpenGenome2's own valid shards. Evo 2 held those out by genomic region, not by species: 31 of its 59 species (80% of its tagged bases) also occur in train, and 91% of it is eukaryotic genomes.
• Species are distinct S__ taxonomy tags. Only eukaryotic genomes and GTDB carry tags; metagenomes, transcripts and viruses add diversity that is not counted as species.

Composition

Every base of OpenGenome2's training shards had the same chance (0.131%) of being kept, so subset shares match the full corpus:

subset	train bases	share	share in OpenGenome2	species	heldout bases
ncbi_eukaryotic_genomes	8.69B	73.8%	74.5%	12,058	93.8M
metagenomes	1.11B	9.42%	9.15%	–	11.0M
eukaryotic_genic_windows	545M	4.62%	4.50%	–	5.8M
gtdb_v220_stitched	463M	3.93%	3.83%	27,259	4.1M
gtdb_v220_imgpr	461M	3.91%	3.83%	–	4.5M
mrna_splice_promoter	257M	2.18%	2.11%	–	2.7M
mrna	151M	1.28%	1.24%	–	1.5M
imgvr_untagged	46.9M	0.40%	0.39%	–	0.5M
ncrna	42.7M	0.36%	0.35%	–	0.4M
imgpr	8.0M	0.07%	0.06%	–	0.1M
organelle	7.9M	0.07%	0.06%	–	0.1M
promoters	0.3M	<0.01%	<0.01%	–	<0.1M

"–" = the subset carries no species tags. Eukaryotic genomes are slightly under-represented (73.8% vs 74.5%); see Sampling note.

Domains (share of train bases): Eukaryota 73.8% · untagged 22.3% · Bacteria 3.8% · Archaea 0.12%.

Largest phyla:

phylum	share		phylum	share
Chordata	33.0%		Bacteroidota	0.54%
Streptophyta	18.8%		Actinomycetota	0.53%
Arthropoda	15.6%		Cnidaria	0.50%
Mollusca	2.48%		Basidiomycota	0.47%
Pseudomonadota	1.11%		Annelida	0.43%
Ascomycota	0.99%		Echinodermata	0.41%

Statistics

split	A	C	G	T	N	GC	soft-masked	CpG o/e
train	28.9%	21.1%	21.1%	28.9%	0.005%	42.2%	32.5%	0.75
heldout	28.8%	21.2%	21.2%	28.8%	0.006%	42.5%	34.8%	0.76
valid	27.8%	22.2%	22.3%	27.8%	0.001%	44.5%	25.3%	0.94

group (train)	bases	GC	soft-masked	CpG o/e
Eukaryota	8.69B	39.5%	42.1%	0.56
untagged (metagenomes, transcripts, viruses, ...)	2.63B	49.1%	6.6%	1.02
Bacteria	449M	54.2%	–	1.17
Archaea	13.9M	47.3%	–	1.09

• The corpus is AT-rich (GC 42%) because eukaryotic genomes dominate; prokaryotes and metagenomes sit at 50–57% GC.
• CpG is depleted in eukaryotes (observed/expected 0.56, methylation-driven) but not in bacteria (1.17).
• Soft-masking (lowercase = repeat-masked) exists only in eukaryotic data: 42% of eukaryotic bases.
• Strand symmetry holds (Chargaff's second rule): 4-mer and reverse-complement frequencies differ by 0.02% on average in train.
• Most frequent 8-mers: AAAAAAAA/TTTTTTTT (0.075% each), then ATATATAT/TATATATA and (CA)n/(TG)n repeats.

Per-subset and per-domain statistics and the full 8-mer counts are in stats/.

Baselines

Order-k Markov models fitted on train (8-mer counts, add-½ smoothing), in bits per base:

split	k=0	k=1	k=3	k=5	k=7
heldout	1.984	1.973	1.961	1.952	1.940
valid	1.993	1.988	1.976	1.969	1.959
train (in sample)	1.982	1.971	1.960	1.953	1.942

Per subset on heldout, best order (k=7 unless noted):

subset	bits/base		subset	bits/base
ncbi_eukaryotic_genomes	1.930		mrna_splice_promoter	1.950
eukaryotic_genic_windows	1.931		mrna	1.967
gtdb_v220_stitched	1.982		ncrna	1.973
metagenomes	1.984		imgvr_untagged	1.980
gtdb_v220_imgpr	1.986		organelle	1.963 (k=1)

• Scored positions are those whose target and 7 preceding bases are all A/C/G/T, identical for every order.
• One model is fitted on the pooled (eukaryote-dominated) composition. GC-rich subsets therefore score near 2 bits, and above 2 at low orders.
• Compare a trained model's masked bits/base on heldout with 1.940.

Format

README.md
plan_train_eukaryotic.json, plan_train_rest.json, plan_valid.json   sampling parameters of each pass
train/<subset>/<shard>.npy     kept windows of one OpenGenome2 shard, concatenated (uint8)
train/<subset>/<shard>.jsonl   one line per window: offset, len, u, tag, rec
train/<subset>/<shard>.done    per-shard statistics
train/selected.json            window selection;  train/manifest.json  composition
heldout/, valid/               same layout
stats/                         statistics (<split>.json), 8-mer counts (<split>_kmer8.npz), baselines (baseline.json)
scripts/                       build, check and load (below)

Sizes: train 12 GB, heldout 129 MB, valid 97 MB; 5.6k files.

• Encoding: one byte per base. Bits 0–2 hold the base (A=0, C=1, G=2, T=3, N or other IUPAC=4), and bit 7 (0x80) marks a soft-masked (lowercase) base. codes & 7 gives the base and codes >> 7 the mask.
• Windows never cross a record, contig or taxonomy-tag boundary.
• tag is the window's full lineage, e.g. D__EUKARYOTA;P__CHORDATA;C__MAMMALIA;O__PRIMATES;F__HOMINIDAE;G__HOMO;S__HOMO SAPIENS, or empty for untagged subsets.

Scripts

script	purpose
og2dataloader.py	PyTorch datasets: RandomCrops (training) and EvalTiles (evaluation, every target once)
og2load.py	numpy reader: iterate windows, filter by subset or taxon, random crops
og2subset.py	the streaming sampler (sample, finalize, report)
build.sh	exact build recipe
og2stats.py	composition, GC, soft-masking, CpG, k-mer statistics
og2baseline.py	Markov baselines
check_rates.py	realized sampling rate per subset
species_overlap.py	species shared between splits

Data loader. Each sample is ctx + 1 consecutive bases of one window.

• Fields: input_ids (base ids, N = 4), labels (the next base), loss_mask and lowercase.
• With mask_non_acgt=True (the default), targets that are not A/C/G/T get label −100 and are excluded from the loss.
• RandomCrops weights windows by their number of possible crops, so every base is equally likely to be seen. It is reproducible per (seed, worker) and can reverse-complement samples (rc_prob).
• Both datasets accept subset= and taxon= filters, e.g. taxon=r"S__HOMO SAPIENS" or taxon=r"D__ARCHAEA".

How it was built

1. Streaming. All 902 training shards of OpenGenome2 (json/*/<subset>/*_train_*.jsonl.gz, 2.79 TB) were streamed and decompressed in memory; only the sample was written.
2. Windows. Each record is split at taxonomy tags (|D__…;S__…|), contig separators (#) and window separators (@). These characters never enter a sequence. Each segment is tiled into 16,384-bp windows on a fixed grid.
3. Per-base-uniform sampling. Each window is kept when a uniform draw u is below the rate (0.131%), whatever its length, so every base has the same inclusion probability. u is stored: og2subset.py finalize --target-tokens N shrinks the sample uniformly.
4. Held-out species. A record goes to heldout when blake2b(species name) / 2^64 < 0.01; untagged records are hashed by record. All records of a species go the same way in every shard and subset.
5. Determinism. Seed 0 and a pinned source revision. Each shard's generator is seeded by (seed, shard path), each record's by (shard, record index), so the output does not depend on worker count or order.

Sampling note

ncbi_eukaryotic_genomes was sampled in an earlier pass. That pass kept a window of length L with probability rate × L / 16,384, which under-samples short windows. Eukaryotic records are ~20 Mb, so almost all windows are full length, and the realized rate is 0.964 × rate. The deficit falls on segment ends next to tags and contig breaks. All other subsets were re-sampled with the per-base rule; their realized rates are 0.995–1.005 × rate (the three smallest subsets: 0.94–1.09, sampling noise). build.sh reproduces both passes exactly, and check_rates.py reports the realized rates.

Reproduce

pip install -r scripts/requirements.txt
bash scripts/build.sh          # stages: euk rest valid finalize stats baseline

About 3–4 hours on 14 CPU cores with ~250 MB/s of bandwidth (CPU-bound). Every stage is resumable.

Limitations

• Context: windows are at most 16,384 bp, so context beyond 16 kb is not available.
• Overlaps from the source: proportions are OpenGenome2's raw ones, including its overlaps. GTDB genomes appear in two subsets (gtdb_v220_stitched, gtdb_v220_imgpr), and mRNA in two (mrna, mrna_splice_promoter). There is no deduplication beyond OpenGenome2's own. Evo 2 trained on a phase-dependent mixture, not on these proportions.
• Repeats: three quarters of the bases are eukaryotic genomes, which are rich in repeats (42% soft-masked).
• Leakage across heldout: untagged records are held out record by record, so related sequences (for example neighbouring metagenome chunks) can fall on both sides of the split.

Citation

Please cite Evo 2, which introduced OpenGenome2:

@article{brixi2025evo2,
  title   = {Genome modeling and design across all domains of life with Evo 2},
  author  = {Brixi, Garyk and Durrant, Matthew G. and Ku, Jerome and others},
  journal = {bioRxiv},
  year    = {2025},
  doi     = {10.1101/2025.02.18.638918}
}

License

Apache 2.0, the license of OpenGenome2, from which every sequence is sampled. The scripts are released under the same license.