furkanbd/r-pfn-checkpoints

Checkpoints: Joint Learning of Factual and Counter-factual Worlds for Causal PFNs

Every checkpoint we trained for the results of the paper, including the fixed-SCM Do-PFN runs. The training and evaluation code, and how each model was trained and evaluated, are in the GitHub repository.

File	Paper model	Size (MB)	sha256
dopfn2d_joint_step262144.pt	Do-PFN 2D	88	114be65332a11d4e…
dopfn1d_J10_step262144.pt	Do-PFN matched resolution	87	8d3fcbfcaa994a47…
dopfn1d_botharms_step262144.pt	Do-PFN matched label	88	40bc2525da832124…
uwyk2d_graph2d_step50000.pt	UWYK 2D	168	7b269408f652c486…
uwykB_bin_step50000.pt	UWYK-B	135	bea21983cc48a980…
uwykB_config.yaml	UWYK-B config	0	0cdffaa8fbf7aeb7…
uwyk_mr_D_j32_step50000.pt	UWYK matched resolution	132	3a820160565e6bef…
uwyk_mr_D_j32_config.yaml	UWYK matched resolution config	0	5255d0f069d5cdbf…
uwyk_ml_C_botharms_step50000.pt	UWYK matched label	135	1b3b9b98ddfa4252…
uwyk_ml_C_botharms_config.yaml	UWYK matched label config	0	0cdffaa8fbf7aeb7…
cpfnC_1d_j1024_step50000.pt	CausalPFN-C	226	01c719bf5081d47f…
cpfnC_2d_j32_eta0_step50000.pt	CausalPFN-C 2D	226	493366de6dc7ff8b…
cpfnC_mr_1d_j32_step50000.pt	CausalPFN-C matched resolution	217	9a8103195d722c84…
cpfnC_ml_1d_botharms_step50000.pt	CausalPFN-C matched label	226	cdda75ce0a70bbb0…
dopfn_backbone_init_s0.pt	shared backbone init (Do-PFN 262k + fixed-SCM)	29	e8899ac05f6f7a23…
fscm10k_natural_step10000.pt	fixed-SCM natural, 10k run (final)	88	c56ca8c1355c2343…
fscm300k_natural_step300000.pt	fixed-SCM natural, 300k run (most recent: step 300000)	88	37ac399b14350c9c…
fscm10k_setup1_step10000.pt	fixed-SCM setup1, 10k run (final)	88	b867c0e8b82c5308…
fscm300k_setup1_step300000.pt	fixed-SCM setup1, 300k run (most recent: step 300000)	88	6d791e21730a88dc…
fscm10k_setup2_step10000.pt	fixed-SCM setup2, 10k run (final)	88	e67ae0c9aace5612…
fscm300k_setup2_step300000.pt	fixed-SCM setup2, 300k run (most recent: step 300000)	88	c84204c74dc39edb…
fscm10k_setup3_step10000.pt	fixed-SCM setup3, 10k run (final)	88	2a24a73d41719a45…
fscm300k_setup3_step40000.pt	fixed-SCM setup3, 300k run (most recent: step 40000)	88	81a58a358253340a…

Published models (not re-hosted)

The paper also evaluates published checkpoints, which are available from their original releases:

• Do-PFN (Robertson et al., 2025): https://github.com/jr2021/Do-PFN, file artifacts/model_submitit_0ccc_id_171b69db_epoch_-1.cpkt.
• UWYK (Reuter et al., 2026): https://github.com/ArikReuter/Graphs4CausalFoundationModels, branch reproduce-realcause-results (commit c27fba6, via git lfs pull), file experiments/checkpoints/full_conditioned_model/final_earlytest_full_conditioning_16773252.0/best_model.pt with its best_model_config.yaml.
• CausalPFN (Balazadeh et al., 2025), the published two-stage checkpoint: causalpfn_v0.pt from https://huggingface.co/vdblm/causalpfn (code: https://github.com/vdblm/CausalPFN).

Their sha256 sums, as used in the paper, are in MANIFEST.txt, so a downloaded copy can be checked against them. Full sha256 sums and the original cluster paths of all files are in MANIFEST.txt.