bench-labs/cagliostro-v2

tokenizer: restrict model_input_names so generate(**tok(...)) works

card: add leaderboard, component and ablation charts

add ablations.png

add components.png

add leaderboard.png

add README.md

add loss_curve.png

add index_curve.png

model card: three-way comparison on one harness

cagliostro-v2: 150M params, 50B tokens, Index 19.99

initial commit