We are getting shortened reply from GPT 4.0, please advice how to overcome this

Hi, everyone.

We are using model gpt-4 (and we are getting the same with gpt-4-32k-0613) through an API to translate some text from English to Ukrainian language.

We’re sending a request with the next prompt:

Translate text into Ukrainian language: Harry Potter has never heard of Hogwarts when the letters start dropping on the doormat at number four, Privet Drive. Addressed in green ink on yellowish parchment with a purple seal, they are swiftly confiscated by his grisly aunt and uncle. Then, on Harry’s eleventh birthday, a great beetle-eyed giant of a man called Rubeus Hagrid bursts in with some astonishing news: Harry Potter is a wizard, and he has a place at Hogwarts School of Witchcraft and Wizardry.

And each time we got a response:

\n\nГаррі Поттер ніколи не чув про Хогвартс, коли листи починають прилітати на килим біля дверей у номері чотири на вулиці Привіт. Але швидко забираються його злим тіткою та дядьком, адресовані зеленою чорнилом на жовтуватому пергаменті з фіолетовою печаткою. А потім, в день його одинадцятого дня народження, у кімнату вривається велетенський чоловік з кулеметними очима, звати якого Рубеус Хагрід, з дивовижною новиною: Гаррі Поттер — це чакл

The translated text is shortened.

When we’re using this prompt:

I have a text. Please translate the text into Ukrainian language: …

We’re getting even shorter text.

We double-checked the amount of tokens and it seems we’re not hitting the limit.

Please guide us in the right direction.

1 Like

My initial instinct is that this is because you’re attempting to translate a copyrighted work so it is limiting the amount of output it produces to avoid adding fuel to the copyright fire.

This appears to be intended behavior and not a bug or something you will be able to work around without violating OpenAI’s usage policies and service terms.

But we saw the same behavior with different texts … so the copyrighted text it seems not the issue here

That may be, I didn’t know, I only saw the text you provided—which is from Harry Potter—and the behavior matches what I’ve seen with other copyrighted inputs—the model stops abruptly.

If you have another, less problematic, example that might be informative.

this text I created and based on it prompt:

I have a text. Please translate the text into Ukrainian language. Text: This smart app is designed to run stand-up meetings with your remote or hybrid team or help you be well-prepared for your classic real-time face-to-face stand-up meetings. The essence of the stand-up meeting smart app is to assist you in doing a stand-up meeting by walking you through your tasks from the task management software in your work chat and helping you categorize them according to classic stand-up meeting questions, such as: What was done yesterday What are you working on now (or today) Tasks which you need help with Once you are done, you get a prepared list of tasks that serves as preparation for the stand-up meeting or is a result of the stand-up meeting. In addition, the smart app will add to the category “:white_check_mark: What was done yesterday” all tasks that were completed between this stand-up and the previous one.

this is the answer:

Ця розумна програма призначена для проведення стенд-ап зустрічей з вашою віддаленою або гібридною командою або допомагає вам підготуватися до класичних зустрічей “в очі в очі” в реальному часі. Суть стенд-ап зустрічі у розумній програмі полягає в тому, щоб допомогти вам провести зустріч, направляючи вас через ваші завдання з програмного забезпечення управління задачами в робочому чаті та допомагаючи вам категоризувати їх згідно з класичними питаннями для стенд-ап зустрічі, так

which is shortened

Magic happened

it was our glitch, we added

‘max_tokens’ => 250,

Sorry :slight_smile: you are most helpful !!!

1 Like



I like to think that if it wasn’t 4am here I would have caught that.

I’m glad you found your solution, I wish you the best of luck on your project!

1 Like

hehe, thanks mate
appreciate your help

This topic was automatically closed 7 days after the last reply. New replies are no longer allowed.