ANTIMEME SYSTEMWORKS
RESEARCH DIVISION
DATASET AUDIT // V1.0_BUILD
NT-REVISED
All 18 tasks hold chr20/21 out of train and 0 of 38,822 test rows reuse a train sequence, reverse complement, locus or interval within their task. One near pair reaches 0.98 identity.
DOC: AM-DA-2026 // NT-REVISED
REV 851F9946252E90C665CDB3CC3EEDB78F1F26197C
18 TASKS // 532,064 ROWS
b4ph/nt-revised-audit
01
Coverage
Tasks
18
ALL FOLDERS
Rows
532,064
493,242 TRAIN // 38,822 TEST
Coordinates parsed
100%
532,064 OF 532,064
Chromosome intersection
0/18
TEST IS CHR20/21 ONLY
02
Per task
Within-task test rows
COUNTS // TEST ROWS| Task | Train | Test | Exactseq | Locus | Overlap | Near≥.98 | Near≥.95 | Unionreport | Chr∩ |
|---|---|---|---|---|---|---|---|---|---|
| H2AFZ | 30,000 | 3,000 | 0 | 0 | 0 | 0 | 8 | 0 | 0 |
| H3K27ac | 30,000 | 1,616 | 0 | 0 | 0 | 0 | 4 | 0 | 0 |
| H3K27me3 | 30,000 | 3,000 | 0 | 0 | 0 | 0 | 6 | 0 | 0 |
| H3K36me3 | 30,000 | 3,000 | 0 | 0 | 0 | 0 | 8 | 0 | 0 |
| H3K4me1 | 30,000 | 3,000 | 0 | 0 | 0 | 0 | 12 | 0 | 0 |
| H3K4me2 | 30,000 | 2,138 | 0 | 0 | 0 | 0 | 4 | 0 | 0 |
| H3K4me3 | 17,468 | 776 | 0 | 0 | 0 | 0 | 1 | 0 | 0 |
| H3K9ac | 23,274 | 1,004 | 0 | 0 | 0 | 0 | 3 | 0 | 0 |
| H3K9me3 | 27,438 | 850 | 0 | 0 | 0 | 0 | 6 | 0 | 0 |
| H4K20me1 | 30,000 | 2,270 | 0 | 0 | 0 | 0 | 5 | 0 | 0 |
| enhancers | 30,000 | 3,000 | 0 | 0 | 0 | 0 | 13 | 0 | 0 |
| enhancers_types | 30,000 | 3,000 | 0 | 0 | 0 | 0 | 13 | 0 | 0 |
| promoter_all | 30,000 | 1,584 | 0 | 0 | 0 | 1 | 3 | 1 | 0 |
| promoter_no_tata | 30,000 | 1,372 | 0 | 0 | 0 | 0 | 1 | 0 | 0 |
| promoter_tata | 5,062 | 212 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
| splice_sites_acceptors | 30,000 | 3,000 | 0 | 0 | 0 | 0 | 5 | 0 | 0 |
| splice_sites_all | 30,000 | 3,000 | 0 | 0 | 0 | 0 | 7 | 0 | 0 |
| splice_sites_donors | 30,000 | 3,000 | 0 | 0 | 0 | 0 | 14 | 0 | 0 |
03
Finding
The README total (544,596) and the Hub listing (532,064) differ by exactly the H3K4me3 train deficit. Test counts match everywhere, and that file is internally balanced at 8,734 rows per class.
04
Method
- Source
- InstaDeepAI/nucleotide_transformer_downstream_tasks_revised
- Revision
- 851f9946252e90c665cdb3cc3eedb78f1f26197c
- Identity
- SHA-256 over canonical sequences
- Loci
- strict name parse, interval tree per chromosome
- Near tier
- bottom-32 k-mer sketch, edlib infix alignment
- Checks
- 28 tests, DuckDB 90 of 90
Reproduce
# from the dataset package
$ python src/acquire.py --work /workspace/nt-audit --pkg .
$ python src/preflight.py --work /workspace/nt-audit --pkg .
$ cat tasks.txt | xargs -P 6 -I{} python src/match_task.py --task {}
$ python src/aggregate.py && python src/verify.py
$ pytest tests -q