Datasets

The annotation corpora gffbase is tested and benchmarked against, where to get them, and what each one is useful for. None is bundled: they are large, they belong to their publishers, and they update on their own schedules.

Everything here is fetched by one script:

python benchmarks/download_corpora.py            # all five, ~257 MB
python benchmarks/download_corpora.py --only mane --only chess   # just two

Downloads land in benchmarks/data/, write through a .part file, and are accepted only after expected size, SHA-256, and gzip integrity checks. An existing file is verified rather than silently skipped.

---

At a glance

Corpus	Format	Feature lines	Download	Fetch with
GENCODE v49 basic	GTF	6,068,892	67 MB	--only gencode-gtf
GENCODE v49 basic	GFF3	6,066,054	85 MB	--only gencode-gff3
RefSeq GRCh38.p14	GFF3	4,932,571	75 MB	--only refseq
CHESS 3.1.3	GFF3	2,761,061	20 MB	--only chess
MANE v1.5 (Ensembl IDs)	GFF3	524,834	10 MB	--only mane

GENCODE ships the same biological release in both GTF and GFF3, which is why both are here. They are related biological releases, but their row models and default inference behavior still differ, so they are not by themselves a controlled measurement of format cost — the published GTF row is measured with inference disabled on both sides, and the GFF3 row with each engine’s defaults. Numbers: Performance. Method: Methodology.

---

What each one is for

GENCODE v49 — the reference human annotation and flagship ingest benchmark. This GTF contains 86,369 gene rows and 293,203 transcript rows; it is not a leaf-only file. The benchmark measures it once with default compatibility inference and once with inference disabled. A third, reproducibly transformed copy removes those parent rows to measure actual synthesis in isolation.

RefSeq GRCh38.p14 — NCBI’s annotation, with Dbxref, Note and gbkey attributes and NC_000001.11-style sequence names. The stress test for attribute handling.

CHESS 3.1.3 — dense relations and custom attributes, across 316 sequences including alt contigs and scaffolds.

MANE v1.5 — one representative transcript per gene, with Ensembl IDs. Small and tidy: the corpus to reach for when you want a real annotation that ingests in twenty seconds.

Important

Three of the five need a duplicate-ID policy

RefSeq, MANE and GENCODE’s GFF3 edition all use the split-CDS convention: one CDS spread over several lines that share an ID. merge_strategy defaults to "error" — as it does in gffutils, which raises on these same files — so you say which reading you want:

create_db(path, "out.duckdb", merge_strategy="create_unique")  # renamed rows
create_db(path, "out.duckdb", mode="strict")                   # one discontinuous feature

See Compatibility & strict modes.

---

Sources

Each is downloaded from its publisher, never re-hosted:

Corpus	Source
GENCODE v49	ftp.ebi.ac.uk — GENCODE, EMBL-EBI
RefSeq GRCh38.p14	ftp.ncbi.nlm.nih.gov — NCBI
MANE v1.5	ftp.ncbi.nlm.nih.gov — NCBI / EMBL-EBI
CHESS 3.1.3	github.com/chess-genome/chess — Salzberg lab

Licensing. Each corpus carries its publisher’s terms, which gffbase does not alter and cannot grant. Cite the annotation you used alongside gffbase — see Citation.

---

Test fixtures

Separately from the corpora above, the test suite uses small vendored fixtures, committed to the repository:

Fixture	What it is
tests/data/*.gff3, *.gtf	Hand-written files covering the hierarchy, GTF synthesis and coordinate edge cases
tests/data/upstream/	32 files copied verbatim from the gffutils test corpus, with PROVENANCE.md recording the commit and licence
tests/data/v1/	A schema-v1 database, gzipped, for the migration tests

These total under a megabyte and need no download, which is why the default pytest run needs no network. The corpus tests are opt-in:

python benchmarks/download_corpora.py
pytest -m corpus

They ingest each real annotation and run the full structural validator over it — the strongest end-to-end correctness evidence in the project.