amino_acid_counts dict | counts dict | created_utc string | dataset_name string | description string | filters dict | length_distribution dict | output_bytes int64 | output_file string | output_sha256 string | source_fasta string | source_sha256 string |
|---|---|---|---|---|---|---|---|---|---|---|---|
{
"A": 5815579,
"C": 1500850,
"D": 3343400,
"E": 4185569,
"F": 3474724,
"G": 4845379,
"H": 1785930,
"I": 4800530,
"K": 4652096,
"L": 7532011,
"M": 3155363,
"N": 3084064,
"P": 3468676,
"Q": 2706277,
"R": 4917311,
"S": 5404899,
"T": 4066856,
"V": 4879442,
"W": 1054910,
"Y": 2249592... | {
"accepted": 1937384,
"noncanonical": 13009,
"records_read": 1950393,
"residue_count": 76923458
} | 2026-09-26T12:49:44.792936+00:00 | uniref50 | Canonical UniRef50 representatives of length 1-50 for peptide MLM pretraining | {
"allowed_amino_acids": "ACDEFGHIKLMNPQRSTVWY",
"deduplicate": false,
"max_length": 50,
"min_length": 1
} | {
"11": 4800,
"12": 5486,
"13": 5506,
"14": 6076,
"15": 6917,
"16": 7489,
"17": 8108,
"18": 8585,
"19": 9844,
"20": 11957,
"21": 13037,
"22": 14646,
"23": 15627,
"24": 16756,
"25": 17833,
"26": 18850,
"27": 20405,
"28": 20503,
"29": 37126,
"30": 42501,
"31": 44222,
"32": 4498... | 78,860,842 | uniref50.txt | 6da9fdbd1519eb9fb8ec61974817552f49f05ce279c0cb7aefc9a3f51a0958f0 | /home/dataset-local/jiahui/pepbenchmark/PepBenchmark/datasets2/uniport/uniref_length_TO_50_AND_identity_0_5_2025_08_28.fasta | 97dc59fb037ad50a37b3871d3219c53412510f661e5cd76d3865cff5c64ad416 |
PepBenchmark UniRef50 peptide pretraining corpus
This dataset contains UniRef50 representative sequences prepared for masked
language-model continued pretraining in PepBenchmark. Each line of
uniref50.txt is one upper-case amino-acid sequence. The file has no header.
Important version note
This is a reproducible reconstruction from the locally archived UniRef50 FASTA snapshot downloaded on 2025-08-28. It is suitable for new training runs, but it is not byte-identical to the unavailable historical corpus used by the original experiment.
| Corpus | Sequences | Residues | Length range |
|---|---|---|---|
| This reconstructed release | 1,937,384 | 76,923,458 | 11-50 |
| Historical experiment | 1,932,360 | 76,692,615 | 11-50 |
Results from a new run should therefore be described as using the reconstructed 2025-08-28 corpus, rather than as an exact reproduction of the historical run.
Construction
The archived source contains 1,950,393 UniRef50 FASTA records. Construction:
- concatenate multiline FASTA sequence records;
- remove whitespace and normalize sequences to upper case;
- retain lengths from 1 through 50 amino acids (inclusive);
- retain only the 20 canonical amino acids
ACDEFGHIKLMNPQRSTVWY; - preserve source order and write one sequence per line without a header.
The source already contains only lengths 11-50. Canonical filtering removes 13,009 records and retains 1,937,384. No additional deduplication is applied; the retained sequences were audited and are unique.
Provenance and complete length/amino-acid counts are recorded in
uniref50.metadata.json.
Integrity
uniref50.txt
size: 78,860,842 bytes
sha256: 6da9fdbd1519eb9fb8ec61974817552f49f05ce279c0cb7aefc9a3f51a0958f0
The archived source FASTA has SHA-256:
97dc59fb037ad50a37b3871d3219c53412510f661e5cd76d3865cff5c64ad416
Training preparation
From the peptide-esm repository root:
python preprocessing.py \
--input_file ./pretrain_data/uniref50.txt \
--output_dir ./processed_data/uniref50 \
--validation_split 0.1 \
--seed 42 \
--min_length 1 \
--max_length 50
This creates a deterministic split of 1,743,645 training sequences and 193,739
validation sequences. The model configuration can then read
processed_data/uniref50/train.txt and
processed_data/uniref50/validation.txt.
License and attribution
The underlying UniRef data are provided by the UniProt Consortium under CC BY 4.0. Users should cite UniProt/UniRef as appropriate. The reconstruction metadata does not replace the upstream database citation.
- Downloads last month
- 39