CoolFace
Datasetpublic

marin-dna/evals_mendelian_traits_rag_harness_255_v1

marin-dna/evals_mendelian_traits_rag_harness_255_v1 Retrieval-conditioned, eval-harness-ready Mendelian SNV benchmark. Each row contains seven fully materialized Zoonomia ortholog slots followed by the shared 127-base human prefix, plus separate reference and alternate completions. Every variant has forward and reverse-complement rows. Produced by the commit-pinned issue #402 RAG pipeline. Model scoring needs only this pinned dataset and a model checkpoint; it does not access… See the full description on the dataset page: https://huggingface.co/datasets/marin-dna/evals_mendelian_traits_rag_harness_255_v1.

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face
0likes22downloads
Dataset Card

marin-dna/evalsmendeliantraitsragharness255v1

Retrieval-conditioned, eval-harness-ready Mendelian SNV benchmark. Each row contains seven fully materialized Zoonomia ortholog slots followed by the shared 127-base human prefix, plus separate reference and alternate completions. Every variant has forward and reverse-complement rows.

Produced by the commit-pinned issue #402 RAG pipeline. Model scoring needs only this pinned dataset and a model checkpoint; it does not access the HAL, genomes, or projection Parquets.

Splits

SplitVariantsRows
train16,14032,280
test9,49018,980

Frozen document contract

  • —Species order/version: zoonomia-rag-v1; human is slot 7.
  • —Projection/version: zoonomia-rag-direct-hal-v1.
  • —context has seven complete non-human slots, seven atomic [SEQ] boundaries, and the shared 127-base human prefix (1,919 tokens before BOS).
  • —Each completion is 128 bases, so context + completion is 2,047 tokens before BOS and 2,048 after the BOS/CLS token.
  • —The centered SNV is absolute token index 1920.
  • —Missing or quality-filtered non-human projections are full 255-base N slots.

Every exact unique 255-base human variant window was projected directly with halLiftover --noDupes against the pinned Zoonomia 447-mammalian 2022 v1 HAL. The build applies the canonical projection pipeline's single-chromosome, single-strand, pre-resize length [128, 512], midpoint-resize, bounds, and strand-normalized sequence conventions. It does not approximate the variant window through a conservation-filtered training anchor.

Provenance

  • —Raw Mendelian source: bolinas-dna/evals_mendelian_traits at 4aed58e50c5dea0b878a665007af2ef9e5108e9f (GRCh38, 1-based SNV pos).
  • —Human sequence-materialized source: bolinas-dna/evals_mendelian_traits_harness_255 at 7b92f047f9a36f90e9ac47886afa2a99264ee35c.
  • —Alignment: Zoonomia 447-mammalian 2022 v1 HAL; source leaf Homo_sapiens.
  • —Derived genomic coordinates: 0-based, half-open.

manifest.json records exact counts, revisions, projection settings, and per-species availability.