soldatmat/CZAI_Summer_School-RFdiffusion_weights

RFdiffusion weights mirror

Hugging Face mirror of two RFdiffusion checkpoints, for the CZAI Summer School 2026 "AI pro enzymy" CZAI_Summer_School-RFdiffusion.ipynb demo notebook.

Important: these are NOT byte-identical to the official release

The official checkpoints are hosted on files.ipd.uw.edu (a University of Washington file server). That host has been down (connection timeouts, not a DNS or local-network issue) for long enough that it could not be used to populate this mirror, and it is also a real risk during a live lecture even when it is up — a single slow or overloaded file server being hit by many students at once.

So the two files in this repo were instead sourced from a third-party Hugging Face mirror, OneScience-Group/RFdiffusion, and cross-checked byte-for-byte identical against a second, independent mirror, GlandVergil/RFdiffusion. Both agree with each other, but neither matches the official MD5 checksums published in RosettaCommons/RFdiffusion's own download_models.sh:

File	MD5 found here (and on both mirrors)	Official MD5 (files.ipd.uw.edu)
Base_ckpt.pt	4aa4a27ba280d23541e01860c106c7cc	6f5902ac237024bdd0c176cb93063dc4
ActiveSite_ckpt.pt	0d9f82af03c73011c6fec060bac5b731	5532d2e1f3a4738decd58b19d633b3c3

We do not know exactly why the bytes differ (e.g. a re-export, a different PyTorch/pickle save, a re-training snapshot) — nobody involved in preparing this mirror controls the OneScience-Group or GlandVergil upload pipeline. What we do know, from directly testing these specific files, is below.

Why we're using them anyway

1. The official host is down. As of this repo's last update, files.ipd.uw.edu was unreachable from multiple networks (this preparation machine and GitHub-hosted CI runners), so there was no way to fetch the byte-identical original to compare against or host instead.
2. RFdiffusion is BSD-licensed, and redistribution is explicitly permitted. The LICENSE states this covers "both the source code and model weights referenced for download in the README file" and is "free for both non-profit and for-profit use." Mirroring a functionally equivalent copy of the weights is within that license.
3. These specific checkpoints were functionally verified, not just checksummed against each other:
   • Both load with strict=True into the real RoseTTAFoldModule architecture defined in RFdiffusion's own code, with the exact tensor shapes the model expects and no missing/unexpected keys.
   • No NaN/Inf values anywhere in either state dict.
   • Both were used to actually run RFdiffusion inference on CPU, including an active-site-scaffolding test that fixed hen egg-white lysozyme's Glu35/Asp52 catalytic residues (via ActiveSite_ckpt.pt) and reproduced their geometry with better than ~0.4 Å RMSD after alignment — the level of fidelity the conditioning mechanism is supposed to achieve. A matched unconditioned control (Base_ckpt.pt, same chain length, no motif constraint) landed at ~2.2 Å RMSD at the equivalent positions, i.e. the conditioned run is several times more accurate, exactly as expected if the checkpoint's weights are functioning correctly. See the notebook's Discussion section for the numbers.

This is a pragmatic, lecturer-approved substitution for a live teaching demo, not a claim that these files are byte-identical to the paper's original release. If you need the official, checksum-verified weights for research or reproduction purposes, fetch them directly from files.ipd.uw.edu once it's reachable, or from RosettaCommons' own distribution channel.

Files

File	Purpose	MD5 (this mirror)
Base_ckpt.pt	General-purpose model, used for unconditional backbone generation	4aa4a27ba280d23541e01860c106c7cc
ActiveSite_ckpt.pt	Fine-tuned for scaffolding small functional motifs (e.g. enzyme active sites)	0d9f82af03c73011c6fec060bac5b731

The notebook's download cell checks the MD5 of whatever it downloads against the checksum expected for the source it came from (this mirror vs. the official host), so a corrupted or wrong download is still caught either way.

License / provenance

RFdiffusion (code and model weights) is released by the University of Washington's Institute for Protein Design under the BSD License, which explicitly states: "This copyright and license covers both the source code and model weights referenced for download in the README file" and "is free for both non-profit and for-profit use." See RosettaCommons/RFdiffusion/LICENSE. This mirror redistributes the checkpoints under those same terms.

Please cite the original work if you use these weights:

Watson, J.L., Juergens, D., Bennett, N.R. et al. De novo design of protein structure and function with RFdiffusion. Nature 620, 1089–1100 (2023). https://doi.org/10.1038/s41586-023-06415-8