zatumout/zipoop_32k1885_010_v4

Zipoop teacher shards (v4)

Per-window "teacher" networks that compress 1,885 ENCODE signal tracks (GRCh38) over 32,768 bp windows at 1 bp resolution. 89,438 windows, stored as one <chrom>.npz per chromosome (25 files, 89.3 GiB).

This is the v4 storage encoding of the same models as zatumout/zipoop_32k1885_010 (v3). Decode output is equivalent; v4 is ~30% smaller.

What a window contains

Each window is one independently trained network plus everything needed to reconstruct its signal:

member	meaning
dna	(32768,) uint8, 0–3 = A/C/G/T, 4 = N — the model input
parameter_0…143 + parameter_names	the trained network weights (144 arrays, ~356k parameters, float16)
model_config	layer/width/head configuration
mean, power, gain, offset	per-track target scaling and calibration
fft_indices, fft_packed, fft_scale	sparse FFT base signal (top-54 rfft bins per track, float16 + one float32 scale per track)
final_gain, final_offset	final per-track affine calibration to real units
valid_v2, valid_encoding	missing-data mask, wire format track-rle-lzma-v2
metadata	the 1,885 ENCODE track records (accession, assay, output type, replicates, experiment)
format, hybrid_format, storage_encoding, byte_shuffle	format tags and the byte-shuffle layout index

Reconstruction is signal = irfft(sparse FFT) + DNA-conditioned residual, then × final_gain + final_offset. Missing positions follow the mask; RNA-seq tracks are 0 and valid where the bigWig did not record, other assays stay missing.

Model description

• Input: 32,768 bp one-hot DNA window (configurable length, this release is fixed at 32 kb).
• Architecture: AlphaGenome-inspired U-Net encoder (7 pooling levels, max pooling), 2-layer / 4-head low-resolution Transformer with sinusoidal positions, shared trunk, per-track heads; width 48, ~356k parameters per window. This is a compact window-specific model, not the AlphaGenome architecture and not a genome-wide predictor.
• Objective: per-track Pearson correlation against the real signal, with a maximum-pooling decoder; the sparse FFT base is fitted first and the network fits the residual.
• Training: independent per-window overfitting — every window gets its own weights. There is no shared model across windows and no train/test split inside a window. That is deliberate: this is fixed-dataset compression, not a generalizing predictor.

Quality and known limits

• Each window is trained until the mean per-track Pearson reaches 0.95 (early export) or it hits a 1,000-step cap, with plateau detection from step 500; windows that reach the cap are exported as secondary-quality and marked in the training receipts (not shipped here).
• A 0.90 mean Pearson is the "usable" label, not a guarantee; low-coverage or repeat-heavy windows are worse.
• Pearson-optimal weights do not guarantee absolute amplitude; the per-track affine calibration (final_gain/final_offset, plus per-window scaling) exists for that, and residual amplitude error remains in difficult windows.
• The track panel is a frozen metadata selection (not per-track signal QC); some accessions may have thin or noisy coverage.
• 89,438 windows have archives; the candidate window list hg38-primary-32k.bed has 94,258, so this release is not a complete tiling of the genome.
• Do not treat these networks as a model of gene regulation or as an ENCODE signal predictor. They reconstruct what was measured in that specific window.

Reading it

Use zipoop>=0.2.0, which decodes both v3 and v4:

pip install -U zipoop

from zipoop import Dataset

with Dataset("path/to/this/folder") as ds:
    w = ds.load_window(ds.windows("chr21")[0])
    # w.key, w.dna (L,), w.signal (L, 1885) float32, w.valid (L, 1885) bool

Lower level: ds.read(window) returns the parsed shard, zipoop.Decoder(shard).decode(start, end, tracks) decodes any slice or track subset, zipoop.evaluation has window_pearson, summarize, calibrate, restore_real.

Integrity: zipoop verify path/to/this/folder checks every file's size and sha256 against MANIFEST.tsv; sha256sum <chrom>.npz works too.

Formats: v3 vs v4

archive	format	hybrid_format	storage
v3	pooper-dna-v3	fft-plus-dna-residual-v1	deflate, fft_coefficients complex64
v4	pooper-dna-v4	fft-plus-dna-residual-v2	LZMA, byte-shuffled floats, float16 fft_packed + float32 fft_scale

zipoop 0.1.x reads v3 only and rejects v4. The v4 path is verified against v3: parameters, masks and calibration arrays are bit-identical, the FFT payload differs only by float16 rounding (max relative error ≈ 4.7e-4), and decoded windows reach a per-track Pearson ≥ 0.9999997 against the v3 decode.

Layout

• <chrom>.npz — zip64/LZMA with members <window>/<key>.npy (164 arrays per window), windows in genome order, names <chrom>_<start>_<end>.
• MANIFEST.tsv — chrom, windows, members, bytes, sha256 per file.
• tracks.tsv — the 1,885 tracks (index, ENCODE accession, modality, assay, output type, replicates, experiment). index is the position along the last axis of every signal / valid array.

Changelog

• v4 (2026-10) — byte-shuffled members, LZMA, float16 sparse FFT; 89.3 GiB (was 128.4 GiB). Content-identical decode.
• v3 (2026-10) — first release, 128.4 GiB, deflate + complex64 FFT.

License and attribution

This dataset is released under CC BY 4.0. You may use, copy, modify and redistribute it, including commercially and including for training or distilling models, provided you give credit:

Zipoop teacher shards, by zatumout (2026), derived from ENCODE data. Licensed under CC BY 4.0.

CC BY has no share-alike condition: models trained or distilled from these shards need not be released under any particular license.

The shards are derived from public ENCODE signal tracks. The ENCODE Consortium asks users of ENCODE data to cite the Consortium's most recent integrative publication, reference the ENCODE accession numbers (listed per track in the shard metadata and in the zipoop panel file) and acknowledge the ENCODE Consortium and the production laboratories (see https://www.encodeproject.org/help/citing-encode). Please do so as well. The zipoop Python package is MIT-licensed separately.