Granitelib Adapters vs Decision Models

Commits scanned

Every commit to the three repos that added or changed decision-model results, newest first. Only model-card READMEs changed; the scores here come from each card as it stands after the last commit listed. A card that comes from an open pull request says so in its header, and the PR is listed under its repo.

Which decision models ran where

How the numbers are combined

• Charts put every score on a 0–100 scale. Cards that report 0–1 metrics (guardian-core, requirement-check) are multiplied by 100. Source tables keep each card's own numbers.
• Margin is the Granite adapter's score minus the best decision model on that benchmark, in points. Pick a decision model under Compare to measure against that model instead.
• Click the Granite chip in the legend to change the adapter. It defaults to "Best per benchmark", the highest-scoring adapter on each benchmark, which is how the guardian and requirement-check cards fill their "Best trained adapter" column. Picking a single adapter, such as the 30B aLoRA, shows that adapter everywhere; the hallucination detection card publishes LoRA adapters only, so an aLoRA choice falls back to the LoRA of the same size there and is flagged.
• On guardian-core, decision-model scores come from the card's comparison table and Granite scores from its per-adapter tables, which report OOD Safety and Function Calling to two decimals.
• Factuality Detection's headline metric is F1 on each of its two test sets, ELI5 and Biographies.
• The RAG cards compare openjev/openjev, which they do not mark as hosted, plus three hosted models. The core and guardian cards compare four hosted models, including clef-flash, and not openjev.