# Code for `AlexWortega/dualrate-physics`

Model-defining code so the checkpoints in this repo are reproducible and loadable
without the full project.

- **`dualrate/`** — the DualRate model package. Importable as `dualrate` (put this
  directory on your `PYTHONPATH`), which matches the loading snippet in the model card:
  `from dualrate.config import DualRateConfig`, `from dualrate.model import DualRatePhysicsModel`,
  `from dualrate.normalization import StateNormalizer`. These three plus `expert.py` and
  `semantic.py` use only in-package relative imports, so checkpoint loading works standalone.
- **`scripts/`** — the training / evaluation drivers: `build_dualrate_cache.py`,
  `train_dualrate.py`, `eval_dualrate.py`, `eval_dualrate_plle_exact.py`. These expect the
  full project layout (they import `src.dualrate.*` as well as `src.physics`, `src.data`,
  `src.training`, `src.evaluation`). To run the full data-build → train → eval pipeline,
  clone the complete repository:

  **https://github.com/AlexWortega/dualrate-physics-llm**

The semantic encoder weights (`Qwen/Qwen3.5-0.8B`) are **not** stored here or in the
checkpoints — they are downloaded from the Hub at load time. The checkpoints hold only the
flow expert, projection, readout tokens, config, and state normalizer.