{{template "head" "Runs"}} {{template "nav" "runs"}}

Training Runs

{{$b := .Benchmarks}} {{if not $b.Runs}}

No benchmark data available. InfluxDB data refreshes every 60 seconds.

{{else}} {{range $b.Runs}} {{$rid := .RunID}} {{$mdl := .Model}}

{{$mdl}}

{{.Type}} {{$rid}}
{{/* Summary stats */}}
{{if hasKey $b.Loss $rid}} {{$loss := getLoss $b.Loss $rid}}
Loss Points
{{len $loss}}
val + train
{{end}} {{if hasContentKey $b.Content $rid}} {{$content := getContent $b.Content $rid}}
Content Scores
{{len $content}}
dimension scores
{{end}} {{if hasCapKey $b.Capability $rid}} {{$cap := getCap $b.Capability $rid}}
Capability Tests
{{len $cap}}
benchmark points
{{end}}
{{/* Training Loss Chart */}} {{if hasKey $b.Loss $rid}}

Training Loss Curve

{{lossChart (getLoss $b.Loss $rid)}}
{{end}} {{/* Content Score Chart */}} {{if hasContentKey $b.Content $rid}}

Content Scores by Dimension

{{contentChart (getContent $b.Content $rid)}}
{{end}} {{/* Capability Chart */}} {{if hasCapKey $b.Capability $rid}}

Capability Benchmark

{{capabilityChart (getCap $b.Capability $rid)}}

Category Breakdown

{{categoryBreakdown (getCap $b.Capability $rid) (getCapJudge $b.CapabilityJudge $rid)}}
{{end}}
{{end}} {{end}} {{template "footer"}}