Minerva 540B
Google DeepMind · 2022-06-30
Minerva 540B is the largest Minerva variant, created by continuing training of Google's PaLM 540B decoder-only model on its mathematics-heavy technical corpus. The paper reports a 2,048-token context, 540.35 billion parameters, and 26 billion continued-training tokens, but the last two disclosures are rounded rather than exact counts; evaluation covered MATH, GSM8K, MMLU-STEM, and OCWCourses with single-sample and majority-vote decoding.
Benchmark scores
| Benchmark | Score |
|---|
No benchmark scores recorded.