MapPFN Weights

Pre-trained and fine-tuned checkpoints for MapPFN: Learning Causal Perturbation Maps in Context (Sextro et al., 2026).

Checkpoints

• model.ckpt — Pre-trained on synthetic biological prior (50 dimensions, 400k steps)
• model_finetuned_frangieh.ckpt — Fine-tuned on Frangieh et al. (2021)
• model_finetuned_papalexi.ckpt — Fine-tuned on Papalexi et al. (2021)

All checkpoints share the same MMDiT architecture (~25M parameters) and differ only in training data. See the GitHub repository for inference and fine-tuning code.

Citation

@inproceedings{sextro2026mappfn,
  title     = {Map{PFN}: Learning Causal Perturbation Maps in Context},
  author    = {Marvin Sextro and Weronika K{\l}os and Gabriel Dernbach},
  booktitle = {Advances in Neural Information Processing Systems},
  volume    = {39},
  year      = {2026}
}

Links: Paper | Code | Datasets | Project Page