Perturbation Bench: A Benchmark of Sequence-Level Perturbation Tasks for DNA Models

A benchmark of sequence-level perturbation tasks for evaluating DNA foundation models. It contains pairs of genomic sequences, including tasks like synonymous codon substitution with 20,000 pairs each for human and mouse. The dataset was created by HuggingFaceBio and was last updated on May 19, 2026.

Use Cases

• Benchmarking DNA foundation models based on pairwise discrimination accuracy between original and perturbed sequences.
• Evaluating model sensitivity to synonymous codon substitutions based on the described syn_human and syn_mouse tasks.
• Testing model robustness to structural alterations in genomic sequences as described in the benchmark.

Strengths

• Includes a specific task with 20,000 sequence pairs for both human and mouse.
• Provides a clear evaluation metric: pairwise discrimination accuracy.
• Last updated on May 19, 2026, suggesting recent maintenance.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for the full dataset is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

HuggingFaceBio

Collection Method

Likely compiled for benchmarking purposes; specific gathering method is not detailed.

Freshness

2026-05-19 12:31:51