Architecture
A bird’s-eye view of how the benchmark turns a ground-truth/prediction pair into metrics. Two entry paths converge on a single evaluator, and an explicit configuration selects which of the eight metric groups run.
High-level architecture: two entry paths (in-memory arrays or GFF/GTF files) converge on one evaluator; the configuration selects which of the eight metric groups run.
How to read it
Configuration (
LabelConfig,AnnotationMode,BenchmarkScope,EvalMetrics) sets the label vocabulary and scope and selects which metric groups run. It is referenced by every layer below. See Annotation Modes and Conventions.Two entry paths converge on the same evaluator:
Path A — in-memory integer label arrays via
gene_calling_benchmark.benchmark_from_arrays().Path B — GFF/GTF files via
gene_calling_benchmark.benchmark_from_gff(), which parses the annotations, pairs transcripts, and builds the paired arrays before handing off to the evaluator. See Getting Started.
Evaluator computes the requested metric groups. Each group answers one question about prediction quality; see Metrics Overview for the full list and Metrics for per-metric detail.
Outputs — figures (
plotting/), corpus-level statistics (compute_global_metrics, GFF path only), and optional Weights & Biases logging (see Online W&B Logging).