Ensembits_cr_data

# Ensembits: data, checkpoints and results (NeurIPS 2026 camera-ready) Companion archive of the Ensembits code release (https://github.com/OliverLaboratory/Ensembits_release). It holds everything the paper's numbers need that is not code. `MANIFEST.tsv` lists every file with its part and size. ## Use Unzip at the data root of the code release (`REPO_ROOT`, see the code README, or set `ENSEMBITS_REPO_ROOT`): cd $REPO_ROOT && unzip ensembits_cr_zenodo.zip Paths inside the archive are relative to that root (`data/...`, `output/...`, `camera_ready/...`), as the code expects them. ## Contents | part | what | size | |---|---|---| | checkpoints | Ensembits λ = 10 (the reported model) and the λ sweep and ablation runs (`output/ensembits_bm_*_rv/`: best checkpoint, config, training history, statistics); AminoAseed and ESM3-struct retrained on our corpus (best checkpoints by validation backbone RMSD) | 3.5 GB | | data | backbone and token caches (mdCATH-div, BioEmu-CATH, MISATO, CryptoBench), RMSF labels, residue sets, splits, the cross-validation file, EC and GO labels, dated SIFTS / UniProt / GO / CATH snapshots | 11.0 GB | | proteingym | ESMFold atom14 structures of the 96 ProteinGym assays used (refolding is not bit-reproducible), per-variant and per-assay scores | 7.6 GB | | results | probe results and RMSF predictions, EC / GO / binding metrics, token ANOVA, ablation, CryptoBench and fold-pair outputs, protein-level statistics | 3.7 GB | ## Not included - **PDBbind-derived files** (binding-site labels, affinity labels, ligand MACCS fingerprints and the binding / affinity prediction files that contain those labels) cannot be redistributed. The code release rebuilds them from PDBbind v2020 with the builders listed in its README (byte-identical to the paper's). - **ESM3 structure-encoder outputs** (token caches, codebook and files computed from them) are under the EvolutionaryScale Cambrian Non-Commercial License; regenerate them with the ESM3 weights from EvolutionaryScale (the code release's builders do this). The retrained ESM3-struct (ours) is included. - **Third-party raw data and model weights** (BioEmu-CATH trajectories, MISATO `MD.hdf5`, mdCATH, CryptoBench, ProteinGym CSVs, ESM3 / ESMFold / ESM2, AminoAseed, ProToken): download them from their sources; `EXTERNAL.md` in the code release gives versions, checksums and licences. ## Licences and attribution Our checkpoints, caches and results are released under CC-BY-4.0 (`LICENSES/CC-BY-4.0.txt`). Derived files keep the terms of their sources, all of which allow redistribution: BioEmu-CATH (CDLA-Permissive-2.0, text in `LICENSES/`), mdCATH and MISATO (CC-BY-4.0), CryptoBench, ProteinGym, ESMFold and ESM2 (MIT). Sources, licences and attribution are listed in `THIRD_PARTY_NOTICES.md`. Please cite these datasets alongside the Ensembits paper.