DNA Foundation Model Benchmark

Compare DNA language models across genomics classification tasks.

Metric

Top N variants

5 50

Variant ranking (best at top)

Plot

Drill into model variants

Select model(s) — single model: all variants as rows; multiple: best variant per family

aidoDNA

Aggregate by dataset (show mean per benchmark dataset instead of per task)

Show all variants (instead of best variant per family, when multiple models selected)

Variant comparison (metric per task)

Mean performance per task category

Pooling/layer ablation — single model

Pick one model, then optionally isolate a single config dimension below.

Isolate dimension

Holds the other two dimensions fixed at their best combo, so bars only differ by this one.

Drill into individual tasks

Select category

Filter by dataset

Performance per task within selected category

Each point = one variant (mean metric across all tasks vs mean speed across tasks). Pareto frontier = variants not dominated on both axes (black outline).

X-axis metric

Throughput (seq/s) Embedding Time (s) VRAM (MB)

Log scale on X-axis

Performance vs efficiency tradeoff

Efficiency ranking

Variants sorted by performance — speed and derived efficiency metrics

Variants sorted by performance — speed and derived efficiency metrics
1	2	3

Performance vs model context size (bubble size = task sequence length)

Applies immediately to all charts — no need to click Apply Filters below.

Only fully-evaluated variants (hide variants missing any task for the selected metric)

Currently returns no data — all runs use subsample_train=0.05.

Production runs only (subsample_train IS NULL)

Exclude models

Pooling strategy

Leave empty to include all pooling strategies.

Model params (min)

1934592 1082130432

Model params (max)

1934592 1082130432

Embedding dim (min)

256 2048

Embedding dim (max)

256 2048

Max context size (min bp)

250 163840

Max context size (max bp)

250 163840

Exclude specific run IDs

Exclude run_ids (comma-separated)

Active rows

Preview (first 50 rows)

Preview (first 50 rows)
1	2	3

Select a model and embedding config to audit which runs were kept or dropped by dedup.

Model

Embedding config hash

Head type

Task category

Task

Pooling strategy

Columns to display

kept

run_id

run_at

subsample_train

task_name

mcc_test

accuracy_test

Runs for selection (✓ = kept by dedup, ✗ = dropped)

Runs for selection (✓ = kept by dedup, ✗ = dropped)
1	2	3

Built with Gradio logo