chorus-borzoi

Chorus mirror: Borzoi (Calico Labs)

Mirror of the Borzoi 4-fold ensemble. The flashzoi PyTorch port repos at johahi/borzoi-replicate-{0..3} are the canonical chorus source; this mirror under the chorus org consolidates weight provenance alongside Enformer, Sei, LegNet, and ChromBPNet for resilience.

Source

• Mirror staged from: johahi/borzoi-replicate-0..3 (the flashzoi PyTorch port maintained by Johannes Hingerl).
• Upstream model: Linder, Avsec et al. 2025, Predicting RNA-seq coverage from DNA sequence as a unifying model of gene regulation, Nature Genetics. https://doi.org/10.1101/2023.08.30.555582
• Reference repo: https://github.com/calico/borzoi

License & attribution

The weights were copied from the johahi flashzoi PyTorch port (which in turn ports the original Calico Labs Borzoi TensorFlow checkpoints). Those weights were created by Calico Labs and are the property of Calico Life Sciences LLC.

The model parameters, outputs, and any derivatives thereof remain subject to the original Borzoi release license (Apache 2.0 per https://github.com/calico/borzoi).

The flashzoi PyTorch port (the actual loader code that consumes these files) is also Apache 2.0; chorus's loader at chorus.oracles.borzoi is released under the chorus license. These terms are consistent with the terms for the reference code and the model weights.

Layout

• fold_0/ — fold 0 PyTorch state_dict + config.json
• fold_1/ — fold 1
• fold_2/ — fold 2
• fold_3/ — fold 3

Total: ~6 GB (each fold has both model.safetensors and the legacy pytorch_model.bin for compatibility).

Usage

In chorus ≥ 0.4 the BorzoiOracle fetches the requested fold from this HF repo by default and falls back to johahi's original repos if HF is unreachable.

from huggingface_hub import snapshot_download
local = snapshot_download(repo_id="lucapinello/chorus-borzoi", repo_type="model",
                          allow_patterns=["fold_0/*"])