Reported by @full depot
Bug Report: Fine-tuned GPT-4o model output quality severely degraded since December 16, 2025
`Steps to Reproduce`
Call fine-tuned GPT-4o model (ft:gpt-4o-xxxx) via Chat Completions API
Send the same prompts that worked correctly until December 15, 2025
Observe model responses
`Expected Result`
Model should return properly structured JSON output with all required fields (sentence_id, translation, slashes, grammar_list) populated with meaningful content, as it did consistently before December 16, 2025.
`Actual Result`
Model outputs contain corrupted tokens:
- Random "22" inserted throughout JSON keys and values (e.g., "22words", "22mot22ives", "2222word_idx")
- Type errors: word_idx returns "false" (boolean) instead of integer
- Structure collapse: "words" field returns ";" (string) instead of array
- Nonsensical values: word field contains ";" instead of actual words
This pattern suggests token generation corruption, not just content quality degradation. The model appears to be hallucinating random numeric
`Environment`
api call