#functions. #multi_tool_use from lightly fine tuned gpt 3.5 1106

API
1

i love 1106 and want to use it but every now and then it hallucinates and just responds with #multi_tool_use or functions.(the function it should run)

its weird

2

If you don’t need parallel tool calls, you can employ the use of the functions API method (not tools).

This will result in less confusion, as the tool wrapper won’t be given to AI for it then to use improperly.

3

already giving it the functions api since the tool api doesnt work…it wont even let you upload a jsonl with the tool api

only uses deprecated function api

4

JSONL, a JSON list with a new JSON on every line, as used for fine-tune or batching, is not going to pass the upload validator for the assistants retrieval tool (which parses inside files and performs inspections and rejects them based on the assumed file extension), and just plain isn’t a supported file type – file types which are listed in the API documentation - supported files.

You cannot use a -1106 fine tune in assistants, so I don’t see how this should matter.

That is unrelated to the initial concern, inability to properly use parallel tools by a model that is fine-tuned, when using the chat completions endpoint. The tuning you can provide degrades the quality of this internal tool, because you can’t train on their use, only train for other types of responses.

5

i might have miscommunicated allow me to explain

I currently dont use the assistants frontend

i use chat completions…specifically i toggle between 3.5 0125 and 1106
i train by uploading a jsonl file

for an example i put tool call sin the jsonl like so

{
“role”: “assistant”,
“content”: “”,
“tool_call”: {
“tool_call_id”: “goodbyeWithoutQuestion”,
“params”: {}
},
“weight”: 1
},
{
“role”: “system”,
“content”: “We have ended the conversation”
}

then…due to how openai doesnt support the tools language in fine tuning and we have to use deprecated function calling

i use a script to transform all my tool calls into function calls

if ‘tool_call’ in message:
tool_call = message.pop(‘tool_call’)
function_call = {
‘name’: tool_call[‘tool_call_id’],
‘arguments’: json.dumps(tool_call[‘params’])
}
message[‘function_call’] = function_call

this works for creating a jsonl file with the functions looking like this

{
“role”: “assistant”,
“content”: “”,
“weight”: 1,
“function_call”: {
“name”: “goodbyeWithoutQuestion”,
“arguments”: “{}”
}
},
{
“role”: “system”,
“content”: “We have ended the conversation”
}