Function words missing at end

I apologize from You people here, this is my first post to this forum. This is a messy report.

The problem mentioned is still with the latest version also. See below.

It is 06.00 UTC and wednesday so very probaly the latest #gpt-4o-2024-08-06 has replaced gpt-4o?? I assumed that gpt-4o-2024-08-06 would be new gpt-4o starting in Wednesday, 0:00 UTC time, this probaply is not the case.

Now 15:05 UTC I tested the problem again both in gpt-4o-2024-05-13 and gpt-4o-2024-08-06, in the former the problem does reproduce. Luckily the latter, the latest, does not have this problem, so I consider this problen fixed.
I did note that gpt4-o was still gpt-4o-2024-05-13.

After a approx. 14000 tokens long chat I get text that is missing function words like β€˜and’ β€˜to’.

I know I can change back to the older version if the new version is to blame, havent tried that yet. I am new here at forum, so probaply there is no point to send the text here right?

Small example from the whole response, the start of response is fine, then it forgots function words. I have the whole chat as a file.

### Balancing Confidence with Humility



1 . Evidence-based Decision Making
     Grounding decisions objective data empirical evidence helps balance confidence realism .

2 . Openness Feedback
     Encouraging diverse perspectives constructive criticism fosters more nuanced understanding reduces risk overconfidence .

3 . Continuous Learning Adaptation
     Embracing mindset continuous learning adaptation enables responsiveness evolving contexts enhances likelihood achieving desired outcomes .

Parameters to the all API calls:

https://api.openai.com/v1/chat/completions

temperature= 0.5,    top_p= 0.5,  frequency_penalty= 0.5, presence_penalty = 0.0

In gpt-4o-2024-08-06 there are missing function words, this is nonsense.

While some may feel confident about their understanding objective truths society translating this confidence into certainty about future outcomes more complex task requires acknowledging inherent uncertainties complexities involved instead focusing on adaptability resilience allows individuals organizations navigate unpredictable landscapes effectively achieving meaningful progress amidst ever-changing contexts

You have caused your own problem.

Do not specify freqency penalty or presence penalty unless you understand the impact of this algorithm. Because: the impact of anything other than 0 is to make it so that that a word cannot be used again, even when, and especially when, it is a common part of language.

Hey that fixed it, thank you.

1 Like