Pull by_query_type accuracy and paired predictions.jsonl from Hub so the Space can show where methods diverge on denied slices, not only aggregate macro-F1.
Add ModernBERT Encoder-FT as a separate RFDT comparator panel.
Keep the generation ladder matched-8B-only; surface Encoder-FT in its own chart, summary strip, findings, and implications so it is not mistaken for a ladder rung.
Discover results/*/metrics.json on the experiments dataset so Refresh picks up
rfdt_base/rfdt_lora (and future slots) without another Space code change.
29 days ago
Replace Gapminder demo with Hub-backed Praxa alignment explorer.
