From 7202cfcd1fa2161343b67771f1ec7a71b78c235f Mon Sep 17 00:00:00 2001 From: Farzad Abdolhosseini Date: Tue, 30 Jul 2024 13:49:15 -0700 Subject: [PATCH] Add "without any explanation" to ST prompt (#60) * add "without any explanation" to ST prompt --- ultravox/data/datasets.py | 10 ++++++---- 1 file changed, 6 insertions(+), 4 deletions(-) diff --git a/ultravox/data/datasets.py b/ultravox/data/datasets.py index b76e0c53..94a3e5c4 100644 --- a/ultravox/data/datasets.py +++ b/ultravox/data/datasets.py @@ -847,11 +847,13 @@ class CoVoST2Dataset(VoiceDataset): } # We currently don't use this dataset for training, so mainly the first prompt it ever used. + # The "no explanation" part is important, specially for evaluations, but it's not repeated + # in all prompts to avoid being too repetitive in training. TRANSLATE_PROMPTS = [ - "Translate the following into {target}: <|audio|>", - "Translate the following into {target} language: <|audio|>", - "Please convert the following into {target}.\n<|audio|>", - "Could you translate this to {target} language?\n<|audio|>", + "Translate the following into {target}, without any explanation: <|audio|>", + "Translate the following into {target} language, no explanation needed: <|audio|>", + "Please convert the following into {target}. Be concise.\n<|audio|>", + "Could you translate this to {target} language? No commentary necessary.\n<|audio|>", "Translate the text below to {target}.\n<|audio|>", "Translate the subsequent text into {target} language. <|audio|>", "Can you translate this into the {target} language?\n<|audio|>",