CoolFace
Modelpublic

lucapinello/chorus-chrombpnet-slim

sourceHugging Faceupdated 5mo agoView on Hugging Face
0likes
Model Card

chorus-chrombpnet-slim

Slim mirror of ChromBPNet + BPNet model weights, used by `chorus` to skip the ~100 GB ENCODE-tarball prefetch when only fold-0 bias-corrected predictions are needed.

What's in here

  • —42 ChromBPNet ATAC/DNase models ({ATAC,DNASE}/<canonical_cell_type>/fold_0/model.chrombpnet_nobias.fold_0.<ENCSR>.h5)
  • —Bias-corrected chrombpnet_nobias variant, fold 0 only.
  • —Each h5 is 25,583,536 bytes, identical to the file inside the corresponding ENCODE tarball at https://www.encodeproject.org/files/<ENCFF>/@@download/<ENCFF>.tar.gz.
  • —744 BPNet/CHIP models (BPNet/<BP_BASE_ID>.h5)
  • —JASPARDeepLearning 2026 release, mirrored from `https://mencius.uio.no/JASPAR/JASPARDeepLearning/2026/models/<BPBASEID>/<BPBASEID>_model.h5`.
  • —Each h5 is 563,760 bytes.
  • —`manifest.json` — sha256, size, original source for every file.

Total: 786 h5 files, 1.49 GB (vs ~100 GB for the full ENCODE tarballs). 67× reduction.

What's deliberately missing

  • —Folds 1–4 of every ChromBPNet model. The shipped chorus default is fold 0, and the upstream Kundaje chrombpnet inference pipeline doesn't ensemble folds. If you need folds 1–4 for advanced ensembling, fetch the original ENCODE tarball and extract: https://www.encodeproject.org/files/<ENCFF>/@@download/<ENCFF>.tar.gz
  • —`chrombpnet` (bias-aware) and `bias_scaled` model variants. chorus's old default was the bias-aware chrombpnet; the new default in chorus ≥ 0.3 is chrombpnet_nobias (which is what's mirrored here). If you specifically need the bias-aware variant, set model_type='chrombpnet' in load_pretrained_model(...) — chorus falls back to the ENCODE tarball flow automatically.
  • —Training / log / per-fold tar archives. Keep what you need from ENCODE.

Provenance

ChromBPNet weights are derived from the ENCODE Annotation ChromBPNet-model set (42 annotations as of 2026-04-28). BPNet weights are mirrored verbatim from the JASPARDeepLearning 2026 release. sha256 in `manifest.json` matches the file inside the source tarball byte-for-byte. Verify with the chorus `audits/2026-04-28chrombpnetslimmirror/` report.

Licence

Each underlying model retains the licence terms of its original distribution (ENCODE consortium for ChromBPNet, JASPAR for BPNet). This mirror does not relicense — it only relocates.

🤖 Built and uploaded with the chorus slim-mirror tooling.