File size: 8,753 Bytes
ca895e3 813d9ee f903cd9 ca895e3 813d9ee ca895e3 f903cd9 ca895e3 813d9ee ca895e3 813d9ee ca895e3 813d9ee ca895e3 813d9ee ca895e3 813d9ee ca895e3 813d9ee ca895e3 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 | ---
license: other
pretty_name: OPLM sequence pretraining data
tags:
- biology
- protein
- protein-language-model
- uniref
task_categories:
- fill-mask
size_categories:
- 1B<n<10B
---
# OPLM sequence pretraining data
The protein sequence pretraining set for [OPLM](https://github.com/briney/oplm), the open
protein language model. Two sources, mixed by rows at training time:
- **uniref70**: UniRef90 release 2026_03 clustered at 70% identity with MMseqs2
`linclust`. One row per cluster: the representative plus every UniRef90 member.
Built by `oplm data uniref70`.
- **deepclust70**: the deduplicated [DeepClust](https://github.com/bbuchfink/diamond/wiki/DeepClust-data-files)
sequences clustered with DIAMOND v2.2.5 `linclust` (`--approx-id 70 --member-cover 90`).
One representative per row, globally shuffled. Re-cut to 1,000,000-row shards by `oplm data reshard`.
Every number below is copied from `uniref70/summary.json`, `deepclust70/train/summary.json`
or `deepclust70/returned.json`, except deepclust70 `train/` residues, which were summed over
the shards. The full provenance is in OPLM's
[`docs/data.md`](https://github.com/briney/oplm/blob/main/docs/data.md).
## Layout
```
README.md
manifest.parquet # one row per file: path, rows, bytes, sha256
uniref70/
summary.json # build record
train/shard-00000.parquet # 758 shards, 100,000 rows each (the last has 18,236)
eval_singletons.parquet # 5,000 held-out clusters with 1 member
eval_gte2.parquet # 5,000 held-out clusters with 2+ members
test_singletons.parquet # 5,000
test_gte2.parquet # 5,000
deepclust70/
train/summary.json # re-cut record
train/shard-00000.parquet # 5,354 shards, 1,000,000 rows each (the last has 896,515)
heldout.json # eval/test screen record
returned.json # record of unused held-out reps returned to train/
eval_singletons.parquet # 5,000 held-out singleton-cluster reps
eval_gte2.parquet # 5,000 held-out reps of 2+-member clusters
test_singletons.parquet # 5,000
test_gte2.parquet # 5,000
```
`manifest.parquet` lists every file with its row count, size and sha256. Check downloads
against it.
## Row schema
uniref70 (one row per cluster):
| column | type | meaning |
| --- | --- | --- |
| `sequence_id` | `string` | the representative's UniRef90 id, e.g. `UniRef90_UPI0018932D0E` |
| `sequence` | `string` | the representative's sequence |
| `member_ids` | `list<string>` | UniRef90 ids of every member, representative first |
| `member_sequences` | `list<string>` | their sequences, same order as `member_ids` |
deepclust70 (one row per sequence):
| column | type | meaning |
| --- | --- | --- |
| `sequence_id` | `large_string` | sequence id |
| `sequence` | `large_string` | sequence |
Sequences are raw one-letter amino acids with no special tokens. deepclust70 had
terminal stop codons (`*`) removed before clustering; internal `*` were kept (about 1.8%
of rows contain one).
## Phase 1 and phase 2
OPLM trains in two phases on the same files:
- **Phase 1** reads `sequence_id` and `sequence` only, so it trains on uniref70
representatives.
- **Phase 2** sets `sample_members: true` on the uniref70 source. Each served row then
yields one member, drawn uniformly within the cluster by a hash of (pass seed, shard,
row). Every epoch redraws it.
Evaluation uses the representatives in both phases.
## Counts
| file | rows (clusters) | member rows | representative residues | member residues |
|---|--:|--:|--:|--:|
| `uniref70/train/` | 75,718,236 | 121,461,869 | 27,363,613,505 | 45,532,832,579 |
| `uniref70/eval_singletons.parquet` | 5,000 | 5,000 | 1,749,551 | 1,749,551 |
| `uniref70/eval_gte2.parquet` | 5,000 | 11,119 | 1,973,383 | 4,469,677 |
| `uniref70/test_singletons.parquet` | 5,000 | 5,000 | 1,701,713 | 1,701,713 |
| `uniref70/test_gte2.parquet` | 5,000 | 11,204 | 1,891,172 | 4,295,018 |
| `deepclust70/train/` | 5,353,896,515 | | 710,712,957,639 | |
| `deepclust70/eval_singletons.parquet` | 5,000 | | 486,576 | |
| `deepclust70/eval_gte2.parquet` | 5,000 | | 887,835 | |
| `deepclust70/test_singletons.parquet` | 5,000 | | 488,725 | |
| `deepclust70/test_gte2.parquet` | 5,000 | | 877,340 | |
uniref70: 121,494,192 UniRef90 sequences in, 75,738,236 clusters out. Members per train
cluster: min 1, median 1, mean 1.604, max 2,370; 60,298,749 train clusters are
singletons.
## Eval and test
Eval and test sets are held out of the whole training mix, not just their own source.
Each stratum is a random draw of that source's clusters: singletons, or clusters with 2+
members. Every eval/test sequence is searched with MMseqs2 against deepclust70 `train/`
and against every member of every uniref70 train cluster:
```
--min-seq-id 0.7 -c 0.8 --cov-mode 0 -s 1 --max-seqs 100000 -e 1e6
```
Any hit excludes the example. For uniref70, every member of a held-out cluster is
searched, not just the representative. Searching every uniref70 member, not just the
representatives, keeps eval and test independent in phase 2, when training samples
members.
Held-out candidates not used for eval or test are in `train/`, unless they match an
eval/test sequence under the same search. deepclust70's are shuffled into the last
`train/` shard.
## Rebuild
UniRef90 input: `https://ftp.uniprot.org/pub/databases/uniprot/uniref/uniref90/uniref90.fasta.gz`,
release 2026_03 (02-Sep-2026), 121,494,192 clusters, 32,107,879,019 bytes, sha256
`f6fa665e27d14fa3dadcff39df6db0052f125ffcf40ba7806f0863c1f262b8dc`, downloaded
2026-10-30.
deepclust70 re-cut (rows and global order unchanged; the sha256 of the
`(sequence_id, sequence)` stream is
`dbf6da39210da1dbbe1f1e63794d91f0b94debbe30a556ccca56e612220a44a4` before and after):
```
oplm data reshard <source deepclust70>/train <out>/deepclust70/train --rows-per-shard 1000000
```
uniref70 (needs `mmseqs` on `PATH`; MMseqs2 18.8cc5c was used). The run was:
```
oplm data uniref70 --uniref90 uniref90.fasta.gz --release 2026_03 \
--deepclust70 <out>/deepclust70/train --workdir <work> --out <out>/uniref70 \
--seed 0 --threads 32 --keep-workdir --min-survival 0.04
```
All other options were at their defaults (`--pool 250000 --n-eval 5000 --n-test 5000
--rows-per-shard 100000 --n-buckets 512 --chunk-rows 200000000`). The clustering stage
had already been run on its own, with the same `easy-linclust` arguments plus
`--split-memory-limit 80G`, so the run above picked up its output. To rebuild in one
command, add `--mmseqs-extra=--split-memory-limit --mmseqs-extra=80G`.
deepclust70 eval/test, then the unused held-out reps back to `train/`
(`<held-out dir>` holds deepclust70's original held-out representatives):
```
oplm data deepclust70-heldout <held-out dir> <out>/deepclust70 \
--deepclust70 <out>/deepclust70/train --uniref70 <out>/uniref70/train \
--workdir <work> --threads 32
oplm data deepclust70-return <held-out dir> <out>/deepclust70 \
--uniref70 <out>/uniref70 --workdir <work2> --threads 32 --seed 0
```
Then write the manifest:
```
oplm data manifest <out>
```
## Versions
uniref70 was built with oplm 0.4.1.dev252+ga400dfede. Clustering ran at commit `ea4b33a`
(2026-10-30), the eval screen at `e6f331c` and the write at `a400dfe` (2026-11-01).
The clustering stage's `summary.json` entry was written by hand from that run's record.
The deepclust70 eval/test screen ran at `ed1d75b` (2026-11-02) and the return of unused
held-out reps at `fa6d689` (2026-11-04); MMseqs2 18.8cc5c throughout.
## License and attribution
The two sources carry different terms, so the dataset as a whole is marked `other`.
**uniref70** is derived from UniRef90 (release 2026_03), copyright the UniProt Consortium,
distributed under the [Creative Commons Attribution 4.0 (CC BY 4.0)](https://creativecommons.org/licenses/by/4.0/)
license (see <http://www.uniprot.org/terms>). The uniref70 files are released under CC BY 4.0.
**deepclust70** is derived from the [DeepClust](https://github.com/bbuchfink/diamond/wiki/DeepClust-data-files)
sequence set (`DeepClustParquet`), a compilation of existing public protein sequence
databases. DeepClust states no license of its own; each underlying database keeps its own
terms, and users of the deepclust70 files are responsible for following them.
Cite UniProt: The UniProt Consortium, *UniProt: the Universal Protein Knowledgebase*,
Nucleic Acids Research.
Cite DeepClust: Buchfink B, Barbé É, Ashkenazy H, Reuter K, Kennedy JA, Drost HG,
*Clustering the protein universe of life using DIAMOND DeepClust*, Nature Methods 23,
724–727 (2026). doi:[10.1038/s41592-026-03030-z](https://doi.org/10.1038/s41592-026-03030-z)
|