add strong-RM checkpoint reward panel (fresh rollouts; training RM + roberta-large IMDB + SiEBERT) 6b036aa verified urdshals commited on 3 days ago
add strong-RM checkpoint reward panel (fresh rollouts; training RM + roberta-large IMDB + SiEBERT) 5ea9a81 verified urdshals commited on 3 days ago
merge pythia-410m eps=0.2 sentiment LLC into dashboard 34b28b8 verified urdshals commited on 7 days ago
backfill prompt-independence (completion_diversity) at LLC checkpoints 5a6cc68 verified urdshals commited on 14 days ago
repetition rate: no-space blob completions count as fully repetitive f340c84 verified urdshals commited on Jun 12
responsive layout: stack side panel below plot on normal screens 591bfb1 verified urdshals commited on Jun 12
410m loc=20k entries (74/84 units; 10 NFS casualties pending mop-up) d581676 verified urdshals commited on Jun 11
yaml provenance panel: move above completions (was below the fold) 6fe4267 verified urdshals commited on Jun 11
hover panel: show SGLD job yaml provenance per checkpoint b99d66b verified urdshals commited on Jun 11
first loc=20k op-point cells (70m fleet + 160m sent-neg) bb28c1b verified urdshals commited on Jun 11
loc 10k-30k + lr 1e-6/3e-6 inventory; new loc=20k/30k sweep cells 39ccb5c verified urdshals commited on Jun 11
fix formal+toxic goal text: MORE toxic (weight -1 on neutral-vs-toxic) 404b69a verified urdshals commited on Jun 10
seed facet = real seeds only; training algorithm gets its own facet e9b6e29 verified urdshals commited on Jun 10
model-keyed inventory; 25 runs; beta0 trio split; gaps filled 4216b23 verified urdshals commited on Jun 10
urdshals commited on 
