Corrected T5x large baselines

Updated T5x-large MNLI and SQUAD baselines
google-research · Aug 27, 2023 · 1fa57af · 1fa57af
1 parent 985ff36
commit 1fa57af
Showing 1 changed file with 1 addition and 1 deletion.
diff --git a/docs/usage/gpu-usage.md b/docs/usage/gpu-usage.md
@@ -35,7 +35,7 @@ For our Pile convergence runs, we used a Global batch size of 2304 for XXL and 2
 | size                                    | GPU              | Precision | #GPUs |  TP   | BS / GPU | Sequences/Sec | Seq/Sec/GPU | Est. Walltime | GPU-days | MNLI 2.0 - matched | SQuAD v1.1 (EM/F1) | Convergence Log                                                                              | Config | 
 | ----                                    | ------------     | --------- | ----- | ----- | -------- | ------------- | ----------- | ------------- | -------- |------------------ | ------------------ | ---------------                                                                              | ----   |
 | [T5-v1.1-small](../t5/t5_1_1/small.gin) | A100 80G SXM     | bf16      | 8     | 1     | 256      | ~5712         | 714         | 4.2 days      | 33       | 83.06%             | 78.33 / 86.63      | [log](https://tensorboard.dev/experiment/lWnHal7PRnOLeZuewyWVxQ/#scalars&_smoothingWeight=0) | [pile](../t5/t5_1_1/examples/small_pile_pretrain.gin)
-| [T5-v1.1-large](../t5/t5_1_1/large.gin) | A100 80G SXM     | bf16      | 64    | 1     | 32       | ~4853         | 75.8        | 4.8 days      | 309     | 90.50%             | 87.31 / 94.04      | [log](https://tensorboard.dev/experiment/aOxJBIvTQBeTJ8XGXxaL6Q/#scalars&_smoothingWeight=0) |[pile](../t5/t5_1_1/examples/large_pile_pretrain.gin)
+| [T5-v1.1-large](../t5/t5_1_1/large.gin) | A100 80G SXM     | bf16      | 64    | 1     | 32       | ~4853         | 75.8        | 4.8 days      | 309     | 89.23%             | 86.12 / 93.21      | [log](https://tensorboard.dev/experiment/aOxJBIvTQBeTJ8XGXxaL6Q/#scalars&_smoothingWeight=0) |[pile](../t5/t5_1_1/examples/large_pile_pretrain.gin)
 | [T5-v1.1-xl](../t5/t5_1_1/xl.gin)       | A100 80G SXM     | bf16      | 144   | 1     | 8        | ~3021         | 21.0        | 7.9 days      | 1,133   | N/A(perf test)     | N/A (perf test)  |                |[pile](../t5/t5_1_1/examples/xl_pile_pretrain.gin)
 | [T5-v1.1-xl](../t5/t5_1_1/xl.gin)       | A100 80G SXM     | bf16      | 256   | 1     | 8        | ~4322         | 16.9        | 5.5 days      | 1,408   | 91.15%             | 89.36 / 95.29      | [log](https://tensorboard.dev/experiment/vuRoEYgkRgWiEtbvgxlOqw/#scalars&_smoothingWeight=0) |[pile](../t5/t5_1_1/examples/xl_pile_pretrain.gin)
 | [T5-v1.1-xxl](../t5/t5_1_1/xxl.gin)     | A100 80G SXM     | bf16      | 512   | 8     | 36       | ~1887         | 3.69        | 12.6 days     | 6,431  |N/A(partial run)   | N/A(partial run)   |                  |[pile](../t5/t5_1_1/examples/xxl_pile_pretrain.gin)