WhenceFade/chreode-downstream

Chreode — downstream fine-tuned heads

Fine-tuned Chreode dynamics heads for Weinreb hematopoiesis (§5.1, Table 1) and Veres islet differentiation (§5.2, Table 2) of arXiv:2605.28111.

Each task is shipped at three seeds (0, 1, 2). Each checkpoint was produced by 5000 epochs of fine-tuning on top of the released pretrained backbone WhenceFade/chreode-pretrained.

Files

File	Task	Seed	Size
weinreb_seed0.pt, weinreb_seed1.pt, weinreb_seed2.pt	Weinreb hematopoiesis (d2 → d4, d6)	0, 1, 2	36 MB ea
veres_seed0.pt, veres_seed1.pt, veres_seed2.pt	Veres islet differentiation (t0 → t1…t7)	0, 1, 2	36 MB ea

These files include the dynamics head only (the scVI encoder is frozen and lives in WhenceFade/chreode-pretrained).

How to use

The full evaluation flow is in reproduce/02_weinreb.md and reproduce/03_veres.md of the Chreode GitHub repo. Quick command (after cloning the GitHub repo + downloading these weights into checkpoints/downstream/):

for seed in 0 1 2; do
  PYTHONPATH=src python -m cellworldmodel.script.run_intermediate_eval \
    --method m10 --dataset weinreb_scvi \
    --experiment g2a_m10_wdit_time2vecu_lowfreqcurl_uncertainty_adamw \
    --model-config-checkpoint checkpoints/downstream/weinreb_seed${seed}.pt \
    --init-checkpoint         checkpoints/downstream/weinreb_seed${seed}.pt \
    --epochs 0 --seed ${seed} \
    --output-dir output/reproduce/weinreb_eval_seed${seed}/
done

--epochs 0 skips fine-tuning and reports eval-only Sinkhorn $W_2$.

Expected paper numbers (3-seed mean ± std)

Weinreb hematopoiesis (Table 1)

Day	Chreode (these weights)	Best baseline
d4	1.5133 ± 0.0757	PI-SDE 1.745
d6	1.6884 ± 0.0362	PI-SDE 1.840

Veres islet differentiation (Table 2)

t	Chreode (these weights)
t1	2.4009 ± 0.0658
t4	2.4048 ± 0.1020
t7	2.9132 ± 0.1704
avg	2.6171

(Full curve t1 – t7 in the paper's Table 2.)

Training recipe

	Setting
Initialization	dynamics_dit.pt from WhenceFade/chreode-pretrained
Epochs	5,000
Optimizer	AdamW β=(0.9, 0.95), wd=0.01, lr=3 × 10⁻⁴
Schedule	5% cosine warmup
Batch	512
Loss	MMD + Sinkhorn W₂ + drift + downhill (1 : 1 : 1 : 0.1)
Hardware	1 × A100 per seed
Wall-clock	Weinreb ≈ 1.5 h / seed, Veres ≈ 2 h / seed

License & citation

MIT — see the GitHub repo. Citation block is the same as the pretrained card.

@inproceedings{qiu2026chreode,
  title     = {Chreode: A Cell World Model for One-Step Temporal Dynamics and Perturbation Prediction},
  author    = {Qiu, Mufan and Zheng, Genhui and Xu, Yinuo and Zhang, Ruichen and Ding, Ying and Long, Qi and Chen, Tianlong},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026},
  eprint    = {2605.28111},
  archivePrefix = {arXiv},
}