# fm-tutorial assets

Pretrained Sudoku checkpoints for [fm_tutorial notebook 05](https://github.com/JChemseddine/fm_tutorial/blob/main/notebooks/05_continuous_discrete_sudoku.ipynb).

Three 28M-parameter DiT checkpoints on *Sudoku Extreme*:
- `vmf_d11_p1/checkpoint.pt` — spherical flow matching (vMF), no time conditioning
- `vmf_tc_d11_p1/checkpoint.pt` — same, with time conditioning
- `masked_p1/checkpoint.pt` — masked diffusion baseline (MDLM-style)

Each `.pt` is a torch.save dict with keys `model_state_dict`, `ema_state_dict`, `config`, `step`, and (for vmf) `warp_state`.

Source: *Spherical Flows for Sampling Discrete Distributions* ([arXiv:2605.05629](https://arxiv.org/abs/2605.05629)).