karanxa/saroku-guard

Swap to report-styled results/adjacent tables, drop old dashboard images

Remove old dashboard-style latency image

Remove old dashboard-style benchmark image

Add adjacent.png (report-styled results table)

Add results.png (report-styled results table)

Embed benchmark/latency/flow charts as images, expand tags, trim redundant tables

Add flow.png

Add latency.png

Add benchmark chart image

Remove duplicate stale Evaluation section

Add benchmark comparison, latency breakdown, adjacent-guard sanity check

Rewrite as a production model card: standard sections, third-person register, remove redundant/internal-facing phrasing

Lead with saroku-guard; move base-model attribution to footer

Update to merged multi-task classifier

Update README.md

Rewrite model card: saroku context, honest metrics caveat, limitations

Replace with pdp-gate-b-v3 (DeBERTa-v3 LoRA classifier)

Document v0.5.0 SDK integration: default fast-path + local:saroku-safety classifier id

Update model card: fix pipeline_tag to text-classification, improve SEO tags, update training data description

saroku-safety-0.5b: 98% binary, 78% 9-class — correct training format, full 9-label coverage

Update model card with latest benchmark results

Update to latest model: 98% binary accuracy, 100% safe recall, 64% 9-class

Add v3.7 benchmark bar chart

Update README with v3.7 benchmark results

saroku-safety-0.5b v3.7: 90% binary accuracy, 52% 9-class

initial commit