Modelpublic

soda-research/soda-hier-1.1b-levanter-checkpoints

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

0likes

Model Card

soda-hier-1.1b-levanter-checkpoints

Native Levanter (JAX) trainer checkpoints — parameters + optimizer state, TensorStore/OCDBT format — for the SODA-Hier scale-up, published so training can be resumed or branched without access to the original cluster:

folder	contents
`trunk-step-42346/`	end of the stable phase of trunk run `soda-hier-1b-08d907e0` — the branch point. Resume it to extend the stable phase to a higher budget, or start a new decay leg from it (see `soda_hier_1b_branch1` in `exp_hero.py` for the pattern).
`branch1-step-52345/`	final state of `soda-hier-1b-branch1-53c95cb9` (decay leg complete, LR = 0) — the trained soda-hier-1.1b model.

To resume: check out the soda-extension branch, download a folder, and point the trainer's load_checkpoint_path at it (training entry points: exp_hero.py + launchers/run_train.sh). Note the training data caches and pick manifests are not included — rebuilding the corpus via preprocess_audio.py is required for an exact continuation.

HF-format weights for inference: soda-hier-1.1b.