Finetuned TabPFN models for binary species distribution modeling (SDM). These models were adapted from the pretrained TabPFN foundation model to improve performance on presence/absence classification tasks common in ecology and conservation biology.

Model Variants

File	Description	Epochs	Val ROC-AUC	Val PR-AUC
`tabpfn-sdm-nonspatial.pt`	Standard (non-spatial) train/test evaluation	100	0.747	0.261
`tabpfn-sdm-spatial.pt`	Spatially-separated train/test evaluation	50	0.653	0.144

The non-spatial variant is trained and evaluated using standard random train/test splits. The spatial variant uses 10km buffer-based spatial separation between training and test data, which is more realistic for ecological applications but results in a harder prediction task.

Training Details

Two-Step Finetuning

Each model was finetuned in two steps:

Step 1 (Cross-Validation): Finetuned on CV-generated train/test splits from training data. Learning rate: 1e-5. Goal: adapt TabPFN to general SDM classification patterns.
Step 2 (Benchmark): Further finetuned on actual benchmark train/test splits. Learning rate: 1e-6. Goal: refine for the specific evaluation setup.

Training Data

226 species across 6 geographic regions: AWT (Australia), NSW (New South Wales), CAN (Canada), NZ (New Zealand), SA (South Africa), SWI (Switzerland)
Species split: 62.5% finetuning, 7.5% validation, 30% held-out test
Data from the disdat R package (standardized presence-only and presence-absence datasets)
Features include environmental covariates (7-12 per region) with native categorical variable support

Training Procedure

Base model: TabPFN v2 classifier (Prior-Labs/TabPFN-v2-clf)
Optimizer: AdamW
LR Schedule: OneCycleLR
Training approach: Iterative R-Python epoch loop with smart sub-batching
- R generates cross-validation data (parallelized with 20 cores)
- Python trains one epoch on GPU
- Smart sub-batching: keeps ALL presences (rare, valuable) in every sub-batch, partitions absences across sub-batches
Max training samples per sub-batch: 1,500
Random seed: 32639

Hardware

GPU: NVIDIA RTX 4000 ADA GPU with 20GB VRAM
Peak memory: ~14GB during training
Training time: ~188 hours total (Step 1 + Step 2) for non-spatial; ~178 hours for spatial

Usage

Python

python

import torch
from tabpfn import TabPFNClassifier
from huggingface_hub import hf_hub_download

# Download model
model_path = hf_hub_download(
    repo_id="rdinnager/tabpfn-sdm-finetuned",
    filename="tabpfn-sdm-nonspatial.pt"
)

# Load finetuned model
device = "cuda" if torch.cuda.is_available() else "cpu"
clf = TabPFNClassifier(
    ignore_pretraining_limits=True,
    device=device,
    n_estimators=8,
    random_state=32639
)
clf._initialize_model_variables()

checkpoint = torch.load(model_path, map_location=device, weights_only=False)
clf.models_[0].load_state_dict(checkpoint["model_state_dict"])
clf.models_[0].eval()

# Fit on training data and predict
clf.fit(X_train, y_train)
probs = clf.predict_proba(X_test)[:, 1]  # Probability of presence

R (via reticulate)

The companion R code provides inference wrappers. See the TabPFN-SDM GitHub repository for the full pipeline.

source("R/run_tabpfn_finetuned.R")

result <- run_tabpfn_finetuned_ensemble(
  train_dat = train_data,
  test_dat = test_data,
  model_path = "tabpfn-sdm-nonspatial.pt",
  max_train_size = 1000L,
  n_estimators = 8L
)

# result$test contains y_truth and y_pred columns

Ensemble Inference (Recommended)

For best results, use ensemble inference which matches the training procedure:

Keeps ALL presences in every sub-batch
Partitions absences across sub-batches for 100% data usage
Averages predictions across sub-batches

This is implemented in run_tabpfn_finetuned_ensemble() (R) or can be replicated in Python using the utilities in the GitHub repo.

Checkpoint Contents

Each .pt file is a PyTorch checkpoint dictionary containing:

Key	Description
`model_state_dict`	Finetuned TabPFN model weights
`config`	Training configuration (hyperparameters, data settings)
`history`	Training history (loss, metrics per epoch)
`step1_path`	Path to the Step 1 model used as initialization

Limitations

Models were trained on species from 6 specific geographic regions and may not generalize to other regions without additional finetuning
Performance varies by species; species with very few presences may have lower accuracy
The spatial variant has lower validation metrics due to the inherently harder task of spatial extrapolation
Requires the tabpfn Python package (v2) for inference
Categorical variable handling requires proper setup (see project documentation)

Citation

If you use these models, please cite the accompanying paper:

bibtex

@article{dinnage2026niche,
  title={A Niche in the Machine: The Promise of AI Foundation Models for Species Distribution Modeling},
  author={Dinnage, Russell and Warren, Dan L.},
  year={2026},
  doi={10.32942/X2VQ10},
  journal={EcoEvoRxiv},
  url={https://ecoevorxiv.org/repository/view/11797/}
}

Please also cite the original TabPFN paper: