subset/README.md

evo3 eval subsets — standard and hard

The exact rows the evo3 standard and hard eval suites score, one parquet per dataset. Exported so these subsets can be scored elsewhere without running the evo3 eval or reimplementing its selection logic.

• Exported: 2026-07-24T19:52:15.198474+00:00
• evo3 commit: 48089c9b12ce559086bfa4a0d0f77f332cf1e378
• pandas: 3.0.3

Layout

standard/<dataset>.parquet
hard/<dataset>.parquet
manifest.json      # full per-dataset provenance (source path, seed, sha256, columns)

Contents

suite	dataset	rows	pos	neg	selection	source rows
standard	clinvar	5000	1675	3325	stratified_representative n=5000 seed=1234	200036
standard	splicevar	5000	2979	2021	stratified_representative n=5000 seed=1234	11978
standard	gnomad_balanced	5000	2500	2500	stratified_representative n=5000 seed=1234	11992284
standard	cosmic	2000	183	1817	stratified_representative n=2000 seed=1234	18903
standard	traitgym_mendelian	3380	338	3042	full dataset (no subsampling)	3380
hard	traitgym_complex	6840	1140	5700	matched_group_representative n=None seed=1234	11400
hard	denovodb	4999	2569	2430	stratified_representative n=5000 seed=1234	68669
hard	causal_mpra	5000	1024	3976	stratified_representative n=5000 seed=1234	12256

Columns

Each parquet carries its source dataset's own columns plus what the evo3 eval loader derives:

• _label — 1 = positive, 0 = negative. Use this rather than the raw label column; the per-dataset positive/negative values are in manifest.json.
• chrom — normalized without a chr prefix (1, not chr1).
• strand — present where the dataset declares strand annotations (ClinVar); minus-strand windows are meant to be reverse-complemented before scoring.
• _subset_stratum — audit key naming the stratum a row was drawn from.

Notes

• Selection is seeded and order-invariant: rows are sorted into a canonical (chrom, pos, ref, alt) order before sampling, so the same seed picks the same variants regardless of the source parquet's on-disk row order. Every subset here was built twice and compared on the variant key before being written.
• traitgym_mendelian carries no subset and is exported in full (~3.4k rows).
• traitgym_complex keeps every positive with its matched controls, so its size is n_pos * (1 + negs_per_pos) rather than a round n.
• cosmic keeps all 183 positives and fills the rest of n=2000 with negatives.
• denovodb holds 4999 rows, not 5000: the subset draws 5000 but one row has a null group_tag, which the eval loader drops when deriving _label. The eval scores 4999 too.

Regenerate with:

PYTHONPATH=src python scripts/eval/export_subset_parquets.py \
    --suites standard hard --out /mnt/weka/shared_datasets/nvidia_subset_parquets