When making a request for the gpt-4-vision-preview model using the OpenAI python library I noticed that the finish_reason is null. Instead there is a finish_details object, which seems to be a replacement for finish_reason
The library types do not indicate this new finish_details object. Nor does the API docs indicate this change
I couldn’t replicate it with another model e.g gpt-4 (turbo or otherwise)
So what’s the expectation here, to continue using finish_reason or rely on finish_details?
Also… please try not to make a breaking change e.g type change, in your API schema design without proper API versioning. Ideally for developers you maintain back-compatibility, unless absolutely necessary
Request body I used:
{
"model": "gpt-4-vision-preview",
"messages": [
{
"role": "system",
"content": [
"You are ChatGPT, a large language model trained by OpenAI capable of looking at images.\nCurrent date: 2023-03-05\nKnowledge cutoff: 2022-02"
]
},
{
"role": "user",
"content": [
"Describe this image for me.",
{
"type": "image_url",
"image_url": {
"detail": "low",
"url": <some url>
}
}
]
}
],
"max_tokens": 256
}
Response I get:
{
"id": "chatcmpl-8IHvXLeFj1fiaxKiYPFJpipihiVIL",
"choices": [
{
"finish_reason": null,
"index": 0,
"message": {
"content": <some content output>,
"role": "assistant",
"function_call": null,
"tool_calls": null
},
"finish_details": {
"type": "stop",
"stop": "<|fim_suffix|>"
}
}
],
"created": 1699369507,
"model": "gpt-4-1106-vision-preview",
"object": "chat.completion",
"system_fingerprint": null,
"usage": {
"completion_tokens": 101,
"prompt_tokens": 141,
"total_tokens": 242
}
}