From f75b6bdb1400418e6f82a2e723c36c0bfd853053 Mon Sep 17 00:00:00 2001 From: Edward Mascarenhas Date: Fri, 7 Feb 2025 02:49:11 -0800 Subject: [PATCH] Adjust baselines for lower number of epochs improved perplexity, lower throughput (#1748) --- tests/baselines/falcon_40b.json | 12 ++++++------ 1 file changed, 6 insertions(+), 6 deletions(-) diff --git a/tests/baselines/falcon_40b.json b/tests/baselines/falcon_40b.json index ab040192c6..f4e26f0a03 100644 --- a/tests/baselines/falcon_40b.json +++ b/tests/baselines/falcon_40b.json @@ -7,9 +7,9 @@ "multi_card": { "learning_rate": 4e-4, "train_batch_size": 1, - "perplexity": 4.0893, - "train_runtime": 360, - "train_samples_per_second": 28.162, + "perplexity": 4.0, + "train_runtime": 550, + "train_samples_per_second": 15.0, "extra_arguments": [ "--bf16", "--gradient_accumulation_steps 16", @@ -42,9 +42,9 @@ "multi_card": { "learning_rate": 4e-4, "train_batch_size": 1, - "perplexity": 4.0893, - "train_runtime": 470, - "train_samples_per_second": 28.162, + "perplexity": 1.6, + "train_runtime": 710, + "train_samples_per_second": 15.0, "extra_arguments": [ "--bf16", "--gradient_accumulation_steps 16",