SOTA · Chinchilla-optimal val loss · 1.3B params, 26B tokens

Every architecture rerun from scratch on the same sealed compute substrate with the same tokenizer and data shard.

loading...
loading... click to expand · verify re-checks sigs + inclusion