This browser is no longer supported.
Upgrade to Microsoft Edge to take advantage of the latest features, security updates, and technical support.
![](data:image/svg+xml, %3Csvg xmlns='http://www.w3.org/2000/svg' height='64' class='font-weight-bold' style='font: 600 30.11764705882353px "SegoeUI", Arial' width='64'%3E%3Ccircle fill='hsl(115.19999999999999, 54%, 21%)' cx='32' cy='32' r='32' /%3E%3Ctext x='50%25' y='55%25' dominant-baseline='middle' text-anchor='middle' fill='%23FFF' %3EIM%3C/text%3E%3C/svg%3E)
Hi,
We noticed that the results of the Azure deployed gpt-35-turbo-v0301 model diverge substantially from OpenAI gpt-3.5-turbo-0301.
These are not minor differences, for the same prompt and parameters results are completely different. The Azure deployment is much less predictable, does not follow instructions, seems like an essentially different model.
Could this be an issue with our specific organization deployed version or is it a known issue?
![](data:image/svg+xml, %3Csvg xmlns='http://www.w3.org/2000/svg' height='64' class='font-weight-bold' style='font: 600 30.11764705882353px "SegoeUI", Arial' width='64'%3E%3Ccircle fill='hsl(22.400000000000002, 11%, 12%)' cx='32' cy='32' r='32' /%3E%3Ctext x='50%25' y='55%25' dominant-baseline='middle' text-anchor='middle' fill='%23FFF' %3ERM%3C/text%3E%3C/svg%3E)
@Itai Marks Thanks for the question, The Azure Open AI gpt-35-turbo is equivalent to the gpt-3.5-turbo model from OpenAI. Can you check the parameters difference if any, Here is the document for the ChatGPT.
![](data:image/svg+xml, %3Csvg xmlns='http://www.w3.org/2000/svg' height='64' class='font-weight-bold' style='font: 600 30.11764705882353px "SegoeUI", Arial' width='64'%3E%3Ccircle fill='hsl(195.2, 78%, 28%)' cx='32' cy='32' r='32' /%3E%3Ctext x='50%25' y='55%25' dominant-baseline='middle' text-anchor='middle' fill='%23FFF' %3EST%3C/text%3E%3C/svg%3E)
I am using the exact same prompt and parameters, just changing the llm from OpenAI to AzureChatOpenAI and getting much worse responses.
![](data:image/svg+xml, %3Csvg xmlns='http://www.w3.org/2000/svg' height='64' class='font-weight-bold' style='font: 600 30.11764705882353px "SegoeUI", Arial' width='64'%3E%3Ccircle fill='hsl(316.8, 68%, 40%)' cx='32' cy='32' r='32' /%3E%3Ctext x='50%25' y='55%25' dominant-baseline='middle' text-anchor='middle' fill='%23FFF' %3EBR%3C/text%3E%3C/svg%3E)
im seeing the same thing, esp for function calling. just a radically different response from azure for the same prompt. and a ton of hallucination.
oh and its also about 10x faster than openAI. which is amazing! but unfortunately not useful with the responses being about 10x worse.
![](data:image/svg+xml, %3Csvg xmlns='http://www.w3.org/2000/svg' height='64' class='font-weight-bold' style='font: 600 30.11764705882353px "SegoeUI", Arial' width='64'%3E%3Ccircle fill='hsl(60.8, 88%, 15%)' cx='32' cy='32' r='32' /%3E%3Ctext x='50%25' y='55%25' dominant-baseline='middle' text-anchor='middle' fill='%23FFF' %3EUM%3C/text%3E%3C/svg%3E)
I too am experiencing azure openai results being different. One thing in particular when I ask it to return an answer from a block of text if the question is lower case for some of the words vs upper case then no answer comes back.
Below is an example:
Answer questions from the knowledge below.
Identification Product form : Mixture Product Name : SILVABRITE 100®
Product code : A00000090 1.2. Recommended use and restrictions on use Recommended use : Alloys for brazing/soldering and other metallurgical processes 1.3. Supplier Lucas-Milhaupt, Inc. 5656 South Pennsylvania Ave. Cudahy, WI 53110 - USA T (414)-769-6000 ******@lucasmilhaupt.com - www.Lucasmilhaupt.com 1.4. Emergency telephone number CHEMTREC within the USA and Canada: 1-800-424-9300 Emergency number : CHEMTREC outside the USA and Canada +1 701-741-5970 SE
Find any product names from the knowledge.
If you cannot find any from the knowledge then say you don't know.
Return the answer in a comma separated string.
When I use the above in openai it works pefectly every time even if I tweak "product names" with Product Names.
But with Azure Open AI i get an empty string if product names is not prefixed with 'Mixture'
This is one of many instances where the models are definitely not the same.
![](data:image/svg+xml, %3Csvg xmlns='http://www.w3.org/2000/svg' height='64' class='font-weight-bold' style='font: 600 30.11764705882353px "SegoeUI", Arial' width='64'%3E%3Ccircle fill='hsl(300.8, 11%, 39%)' cx='32' cy='32' r='32' /%3E%3Ctext x='50%25' y='55%25' dominant-baseline='middle' text-anchor='middle' fill='%23FFF' %3EBC%3C/text%3E%3C/svg%3E)
I've found the same thing here, specifically function calling is just unusable in Azure OpenAI. These are not equivalent models, or the function calling must be stuffed into the system prompt of Open AI's API.