FIne-tuned GPT-4o Model Output Format Issues and Timeouts After Recent Service Outage

Hello
We are currently using a fine-tuned GPT-4o model via the OpenAI API

1. Issue Summary

Before the recent OpenAI service outages, this fine-tuned model was responding consistently and reliably, including producing outputs in a strictly defined JSON format as expected.

However, after the outage, we have observed the following issues, despite no changes to our prompts, system messages, or API request parameters:

  • Responses occasionally return undefined, empty, or malformed outputs
  • Outputs sometimes do not follow the expected JSON schema
  • Intermittent API request timeouts when calling the fine-tuned model

2. Example Logs

Below are representative backend logs showing the timeout errors:

[WARNING] FT request failed for sentence s1: OpenAI API request timeout:

[FT MODEL ERROR] Sentence ID: s1
Error: OpenAI API request timeout:

[WARNING] FT request failed for sentence s4: OpenAI API request timeout:

[FT MODEL ERROR] Sentence ID: s4
Error: OpenAI API request timeout:

[WARNING] FT request failed for sentence s8: OpenAI API request timeout:

[FT MODEL ERROR] Sentence ID: s8
Error: OpenAI API request timeout:

3. Key Questions

We would appreciate clarification on the following points:

  1. Can updates, outages, or internal changes to GPT-4o base models affect the behavior, reliability, or output formatting of existing fine-tuned models?
  2. Are there any known issues or ongoing incidents related to fine-tuned GPT-4o models following the recent outages?
  3. Are there recommended mitigations or best practices to:
  • Improve output format stability (e.g., strict JSON adherence)
  • Reduce or handle timeout errors more effectively
  1. Is re-fine-tuning or migrating to a newer base model version recommended in this situation?

4. Additional Notes

  • The correct fine-tuned model name is being passed in all API requests.
  • No changes were made to:
    • Prompt structure
    • Output schema
    • Temperature / max tokens
    • API client implementation

We rely on this model in production, so any guidance or confirmation would be greatly appreciated.

Thank you for your support.

PS. Additional Observation Regarding JSON Output

In addition to the issues described above, we have observed a recurring pattern in the malformed JSON responses:

  • The model occasionally appends the number 22 (or variations including 22) at random positions within the JSON output.
  • This often results in:
    • Invalid JSON syntax
    • Extra numeric tokens appearing after closing braces
    • Corruption of otherwise well-structured JSON responses

This behavior was not present prior to the recent service outage and occurs intermittently under the same prompt and configuration.

example
“105”:22, “words229”:22, “words”:22994, “words”:22, “words”:22, “22”:22, “words”:2222, “words”:22, “words”:22, “words”:22, “words”:22, “words”:22, “words”:22, “words”:22, “words”:22994994, “words”:22, “words”:22, “words”:22994, “words”:22994, “AT”:2222, “words”:22, “AT”:22, “AT2222,37”:22, “words”:22, “AT”:22, “AT”:22, “AT”:22, “AT”:22, “AT”:22, “AT”:22, “AT”:22, “AT”:22, “AT”:22, “AT”:22, “AT”:22, “AT”:22, “AT”:22, “AT”:22, “AT”:22, “AT”:22, “AT”:22994, “words”:22

The system returned to normal operation as of 2025-12-18 00:43 UTC.

4 Likes

We are facing the same issue today, after service outages. No changes on our side. We have fine tuned model based on 4o model too.

2 Likes

We are experiencing the exact same issue since yesterday 2025-12-16 21:00 UTC.

3 Likes

We are also experiencing the same issue. Fine-tuned 4o model that is returning very long, random responses.

Acadia were challenging in terms182 of playing2237 CROSS. I had to learn to be patient39,experimental to trust183 the processexperimentalexperimentalexperimentalexperimentalexperimentalexperimentalexperimentalexperimentalexperimentalexperimentalexperimentalCaraexperimentalexperimentalexperimental,18422experimental22183experimentalexperimental22experimentalexperimental22experimentalexperimentalexperimental and to believeexperimentalexperimentalexperimental in22 myself. It was a periodexperimental2222 of growthexperimental22 that happen

1 Like