Tdig Evo2 Hidden States

Hidden state activations from the Evo2 7B transformer model's forward pass on human genomic sequences. The dataset includes raw and RMS-normalized states for reference chromosomes 22 and 17 (hg38) and 10,910 ClinVar variants across 15 cancer genes. It was created by author darejinn as a companion to the TDiG (Think Deep in Genome) project and was last updated on May 25, 2026.

Use Cases

• Analyzing internal model representations (hidden states) based on genomic sequence inputs.
• Investigating model behavior on known pathogenic variants based on the included ClinVar data.
• Benchmarking interpretability methods for genome-scale transformer models based on the provided hidden states.

Strengths

• Includes data from a specific, named model (Evo2 7B base with 8K context).
• Covers a defined set of 10,910 ClinVar variants in 15 cancer genes.
• Provides both raw and RMS-normalized hidden state values.

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

huggingface

Collection Method

Forward pass of the Evo2 7B transformer model on specified genomic sequences.

Freshness

Last updated 2026-05-25 05:57:51; freshness should be verified.