Context
A neural decoder is only useful if the repository can generate, version, and evaluate its training data reproducibly. The current examples include decoder comparison and a trained JSON weight model, but there is no full training/evaluation pipeline that ties together dataset generation, manifests, model training, and fair benchmarking.
Proposal
Add a reproducible pipeline for neural decoder datasets and evaluation.
Suggested stages:
- generate labeled examples from Surface, hybrid/GKP, CSS, and hardware replay workflows,
- store dataset manifests with code distance, noise model, p/sigma grid, seed ranges, and split identity,
- train at least one baseline model from repo scripts,
- export model artifacts for the runtime backend,
- compare MWPM, UF, BP, neural-guided MWPM, and neural-only decoders using the same seeds and sweep grids.
Acceptance Criteria
- A command/script generates a small deterministic dataset fixture suitable for CI.
- A larger example workflow generates train/validation/test splits with manifest files.
- Training scripts record model config, git commit, seed, data manifest hash, and objective metrics.
- Evaluation emits a single comparison table/CSV for all supported decoders.
- Docs explain how to reproduce the baseline model and how to plug in a custom model.
Context
A neural decoder is only useful if the repository can generate, version, and evaluate its training data reproducibly. The current examples include decoder comparison and a trained JSON weight model, but there is no full training/evaluation pipeline that ties together dataset generation, manifests, model training, and fair benchmarking.
Proposal
Add a reproducible pipeline for neural decoder datasets and evaluation.
Suggested stages:
Acceptance Criteria