Boris-2-0907 and Boris-2-0917
On September 7th, 2026, Boris-2 began training for the first time. Notably, it was lagging severely behind its competitors. At 30 billion tokens, it couldn't break below ~3.100 loss. The ever-intelligent coding agent, Claude Opus 5, said that it was "merely a side effect of WSD learning." It was not. Finally, the bug was discovered: a mismatch between the Muon and AdamW gradients. I discovered this about 10 days into training.
On September 17th, 2026, Boris-2 began training for the second time under the name "Boris-2-0917," in contrast to the previous version, "Boris-2-0907." Once again, the omnipotent and eternally right coding agent led me astray. It reiterated that the old tokenizer, despite being trained on an older version of the corpus, was perfectly fine. It was actually leaving gains on the table that it said "did not matter." The same was true of the n-gram mixture.
Finally, I made the decision to cull the run at around 7 billion tokens. Despite 0917 being perfectly serviceable, and really not too much worse than the new "Boris-2-0920," I decided that no gains should be left on the table.
Today, I am making those former versions of Boris-2 open-weight, free for anybody to use, modify, or distribute.
— KlondikeDev (Joseph Jones)