Jugc/SFCD

Spherical Flows for Sampling Categorical Data — model checkpoints

Trained checkpoints for Spherical Flows for Sampling Categorical Data (Chemseddine, Kornhardt, Steidl). Code, configs, and evaluation scripts: github.com/JChemseddine/SFCD.

One folder per model, <dataset>/<model>/, each holding the training checkpoint final.pt and its exact training configuration config.json. Naming: vmf (von Mises–Fisher path), geo (geodesic/SLERP), vp / ve (Euclidean baselines); _tc = CDCD-style learned time warp; _selfcond_scp025 = self-conditioning (p = 0.25).

Dataset	Models	Training
lm1b/	vmf, geo, vp, ve × plain / _tc / _tc_selfcond_scp025 (12)	1M steps
owt/	vmf (plain/tc/sc), vp + geo (tc/sc) (7)	1M steps
tinygsm/	vmf, geo, vp, ve × _tc / _tc_selfcond_scp025 (8)	250K steps

Usage

from huggingface_hub import hf_hub_download

ckpt = hf_hub_download("Jugc/SFCD", "lm1b/vmf_tc_selfcond_scp025/final.pt")
cfg  = hf_hub_download("Jugc/SFCD", "lm1b/vmf_tc_selfcond_scp025/config.json")

Evaluate with the repository's scripts, e.g.

uv run python src/scripts/eval/eval_lm1b.py \
    --config <cfg> --checkpoint <ckpt> \
    --method pc_ode_warp --sampling-steps 64 --corrector-steps 1 \
    --corrector-interval 1 --corrector-epsilon 1e-3 --corrector-scaling on

See the repository README for data preparation, training, and the full replication grids behind the paper's tables.