Genomic-NIAH: A Long-Context Retrieval Benchmark for DNA Language Models

Genomic-NIAH is a benchmark dataset created by HuggingFaceBio to measure the long-context retrieval capabilities of genomic language models. It plants random (KEY, VALUE) DNA pairs within real-genome sequences and challenges models to retrieve the VALUE given the context and KEY. The dataset was last updated on May 19, 2026.

Use Cases

• Benchmarking long-range retrieval performance of genomic language models based on the described (KEY, VALUE) pair task.
• Evaluating a model's ability to use its full context window on DNA sequences, as the task requires retrieval of uncorrelated planted pairs.
• Testing whether a model memorizes or genuinely retrieves information from long genomic contexts.
• Comparing different model architectures on a standardized genomic retrieval task.

Strengths

• Designed specifically to test a core capability (long-context retrieval) for genomic AI models.
• Based on real-genome sequences, providing a biologically relevant context.
• Task design isolates the retrieval mechanism by using KEY-VALUE pairs uncorrelated with the surrounding sequence.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count 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 constructed by planting synthetic (KEY, VALUE) DNA pairs into real genomic sequences.

Time Range

null

Freshness

Last updated 2026-05-19 11:29:10; freshness should be verified.

Geography

null