Models¶
The models are bundled with the package and selected by name.
The current family¶
The SO3LR Tauri models, described in the SO3LR Tauri preprint, come in three sizes:
so3lr-tauri-m(default): start here; the model we use for biomolecular MD.so3lr-tauri-l: the most accurate, recommended for reactions; 3.7× slower than-m.so3lr-tauri-s: 1.4× faster than-m.
The Tauri models cover H to At (Z = 1–85), and the original so3lr covers H,
C, N, O, F, P, S and Cl. Structures with any other element are rejected with
an error.
Legacy SO3LR model¶
The model from the JACS paper is kept,
as so3lr, to reproduce published results:
Names and versions¶
Names follow so3lr-tauri-<size>[-<release>], where <size> is s, m or
l and <release> is the release date as YYMM:
| Name | Means | Stability |
|---|---|---|
so3lr-tauri-m |
the current M model | moves when a new release lands |
so3lr-tauri-m-2606 |
one release | never changes |
The currently bundled releases are:
| Release | Canonical names |
|---|---|
| 2606 (current, default) | so3lr-tauri-{s,m,l}-2606 |
| 2602 | so3lr-tauri-{s,m,l}-2602 |
| Original | so3lr |
The 2602 models are the models after stage 4 of training, and 2606 the final models after stage 5 (see the preprint).
To pin a release for papers and benchmarks, use its dated name, e.g.
so3lr-tauri-m-2606, because the short names give different numbers after an
upgrade.
Selecting a model¶
Use the same name in Python, on the command line and in settings files:
Custom and fine-tuned models¶
To use a fine-tuned or custom model, pass a path instead of a name:
The directory needs a hyperparameters.json and either a params.pkl or an
orbax checkpoints/ folder, which is what fine-tuning
writes to its --workdir.
Verifying the bundled files¶
The file so3lr/models/SHA256SUMS lists the SHA-256 of every bundled file. To
check them from the repository checkout:
A match shows that the files are identical to the release.