syuekai/gut-baseline-scoring-refs

gut-baseline-scoring reference classifiers

Pinned reference artefacts for the gut-baseline-scoring task in the science-tasks benchmark repository. Not a dataset: two QIIME 2 2026.1 naive-Bayes taxonomy classifiers shipped to the task's environment image.

Both were trained from the pinned SILVA 138.1 NR99 release (silva-138-99-seqs.qza / silva-138-99-tax.qza, sha256-pinned in the task's panel/resources.tsv) with qiime feature-classifier extract-reads + fit-classifier-naive-bayes:

• classifier_v4.qza — V4 (515F/806R) region, primers GTGCCAGCMGCCGCGGTAA / GGACTACHVGGGTWTCTAAT, extract --p-trunc-len 0 --p-min-length 100 --p-max-length 400
• classifier_v3v4.qza — V3–V4 (341F/805R) region, primers CCTACGGGNGGCWGCAG / GACTACHVGGGTATCTAATCC, extract --p-trunc-len 0 --p-min-length 300 --p-max-length 500

Each .qza is a QIIME 2 archive and carries its full action provenance (software versions, parameters, input checksums) inside. The training recipe is also documented in the task's environment/Dockerfile comments.

Training is not re-run inside the task's image build on purpose: the fits hold tens of GB resident each (measured peak for the V4 fit alone: 23 GB) and add hours to every image build; the pinned archives keep the build deterministic and minutes long.