6 days ago
-
Card: BLiMP count note — cause of the 47 913 count not determined
-
Card: BLiMP count 47 912 (one pair differs at a numerical tie on another GPU)
-
Scenario chart: continuation from checkpoint 1050 (before decay); remove resume-from-decayed scenario
-
Add training-loss scenario chart (backtest failed; not a forecast); fix decay shape (1 - sqrt)
-
Card: Polish leaderboard columns (two independent implementations)
-
Card: final training facts and leaderboard-protocol results
-
Card: available vs planned tokens; no empty sections
-
Card: live training curve link
-
Model card: architecture, data, recipe; training in progress
