Metrics Dashboard (/metrics)¶
Aggregate model comparison across held-out genes — the counterpart to the per-gene pages.
Two sources¶
- Base-model runs are discovered from evaluation outputs under
examples/base_layer/output/. - Meta-model runs are registry-driven: the dashboard lists what
settings.yamldeclares undermeta_models, then loads that model's evaluation JSON. It never guesses from directory names, so a run appears here only if it is a registered model.
Models registered status: research (currently M3-S) are excluded. They are evaluated by metrics
this dashboard does not display, so listing them would show empty columns.
Reading a comparison¶
Each meta model is shown against its own base model — the one it refines, from the registry's
base_model field. Comparing M2-S against a base it never saw would be meaningless.
The provenance fields tell you what a model was built from:
| Field | Example | Why it matters |
|---|---|---|
variant |
M2-S |
which task: canonical / alternative / novel |
arch |
concat_fusion |
the network, named for its mechanism |
corpus |
cleanannot |
the data generation |
train_annotation |
ensembl |
which annotation it learned from |
eval_protocol |
alt_sites |
which truth set its headline number uses |
train_annotation is the one people skip and shouldn't. M1-S learned MANE; M2-S learned Ensembl.
A number is only interpretable once you know which.
Which numbers are safe to quote¶
Never quote FP or FN at threshold 0.5
Splice sites are under 0.5% of positions. At a fixed 0.5 cutoff the M2-S genome-wide evaluation shows FN −89.8% and FP 1,006,348 against base 13,163. Both are real; both are artifacts of the operating point. Report PR-AUC (threshold-free) or counts at each model's F1-optimal threshold, and say which you used.
Precision was corrected in August 2026
The evaluator subsamples "neither" positions to 1% to keep PR-AUC tractable over 600M positions. Threshold sweeps had been counting false positives over the retained rows, inflating precision on the FP side by roughly 100×. This dragged every stored F1 optimum far too low — M1-S meta read 0.35 where the corrected value is 0.99.
Recall, TP and FN were never affected, because positives are never subsampled. That is exactly
why it stayed hidden: every recall-side headline was correct. Figures on the results pages are
corrected; older screenshots and any FP 346 figure are not.
Related¶
- Reading the numbers — the general version of the cautions above.
- Published results — the citable held-out figures.
- Reporting & promotion — how a run becomes a registered model.