Model profile

TabICL v2

TabICL v2 is a transformer pretrained on varied synthetic tables for classification and regression. It uses labelled training rows as context when predicting new rows, with attention designed to handle larger tables without requiring equally long tables throughout pretraining.

Tabular FoundationRegular feature limit: 2,000Source ↗

Model profile

TabICL v2

Default parameters. Peer models use their prespecified benchmark-default configuration without per-dataset tuning. This card describes that configuration, not the best achievable tuned model.

Operating map

Macro-F1 Elo relative to Random Forest. Columns are training-sample caps; rows are dataset modalities. Small labels show the number of binding targets.

Elo difference vs. Random Forest

Fixed nonlinear scale; extremes capped at ±1,000 Elo.

No result / comparison unavailable

Reliability

Failure rate map

Dataset modality

Failed fits / (successful + failed fits) for the selected model and modality. Design skips are omitted. A dash means no attempted fits; 0% means all attempted fits succeeded. Hover for counts. Colours span 0–20%, with rates above 20% using the darkest shade; cell labels and hover retain the actual rate. This chart is independent of the performance sensitivity view.

Reference benchmark results

Strict Macro-F1 Elo at p=10,000, n=100. Random Forest is anchored at 1,000; intervals are target-bootstrap 95% intervals. These reference results stay fixed while you explore other budgets in the charts. Overlapping intervals do not establish a clear winner.

Dataset modality	Elo	95% interval	Targets
All modalities	786	684–875	44
DNA methylation	530	-750–690	4
GLM/PLM Embeddings	973	469–1,414	4
Gene expression	498	246–663	10
Genomic prediction	316	133–441	4
Metagenomics	463	280–583	10
Molecular properties	1,080	965–1,180	10
Other	894	-941–1,137	2