Flan-T5 Base
Google · 2022-10-20 · 248M parameters
Flan-T5 Base is the instruction-finetuned T5 v1.1 Base checkpoint from Google's 2022 FLAN scaling work, a 12-layer-per-stack encoder-decoder released as open weights. It applies the same broad instruction tuning, including chain-of-thought data, as the other Flan-T5 checkpoints; the paper rounds its size to 250M parameters, while the released checkpoint holds 247,577,856.
Benchmark scores
| Benchmark | Score |
|---|---|
| BBH (BIG-Bench Hard) | 27.9 |
| MGSM | 0.4 |
| MMLU | 35.9 |