metadata
datasets:
- bigscience/P3
language:
- en
A 3B T5 model trained on the P3 (T0 split) dataset for 20,000 steps with a batch size of 2048 a maximum input sequence length of 1024, a maximum output sequence length of 256, and the Adafactor optimizer with a constant learning rate of 0.001. The model is trained from the T5 v1.1 lm-adapt checkpoint and fully finetuned.
For more details, see HINT: Hypernetwork Instruction Tuning for Efficient Zero- & Few-Shot Generalisation.
Performance on T0 held-out tasks (average accuracy across prompts using rank classification):
| Model | ANLI (avg) | HellaSwag | StoryCloze | CB | COPA | RTE | WiC | WSC | WinoGrande | Average |
|---|---|---|---|---|---|---|---|---|---|---|
| T0-3B | 33.4 | 27.2 | 84.0 | 45.4 | 75.9 | 64.6 | 50.7 | 65.1 | 51.0 | 55.2 |
| hypertask_T0_3B (this model) | 41.7 | 30.1 | 96.9 | 72.7 | 89.1 | 81.2 | 51.7 | 57.2 | 59.2 | 64.4 |