Views
No views yet
CLAUDE.md in the source repo
(https://github.com/gesttaltt/3-adic-ml) cites as evidence for its documented
results (V15–V23). These previously existed only on one local disk with no
backup — .gitignore in the source repo has referenced a
checkpoint_hub download tool since early on, but that tool was never
actually built. This repo is that archive, added retroactively.checkpoint.pt (best_Q.pt from the run), config.yaml
(exact training config), and results.json where the run produced one.| Folder | What it is | Cited in CLAUDE.md as |
|---|---|---|
v15.0_long_term_stability/ | 2000-epoch long-horizon stability run | V15.0 completed |
v16.0_human_fine_tuning/ | Fine-tuned on human TP53 codon windows | V16.0 (Human TP53) completed |
v17.0_rosetta_manifold/ | Grand-master run: synthetic + human TP53 + bioactive peptides, single manifold | Phase 17 "Rosetta Manifold" |
v17.1_rosetta_manifold_resume/ | Resume/continuation of the above | Phase 17 resume |
v21.0_clean_run/ | Best checkpoint epoch 230/1000. ARI(prefix-3, v=0)=1.000, Spearman hierarchy=0.8335, Q=1.9755 | "V21.0 Training Results" |
v23.0_algebraic/ | Best checkpoint epoch 210/1000. Adds algebraic structure signal (4-bit signature, Lagrangian dual ascent). Q=1.9698, hierarchy Spearman=0.8335 | "V23.0 Training Results" |
v17.0_rosetta_manifold has no results.json (wasn't saved at the
time of that run) -- config + checkpoint only.1import torch
2ckpt = torch.load("v21.0_clean_run/checkpoint.pt", weights_only=True)src/models/vae.py (TernaryVAEV6Controllable) for
the model class, and the corresponding config.yaml in each folder for the
exact architecture/training hyperparameters each checkpoint was produced
with.