The GPT-4-1106-preview model keeps generating "\\n\\n\\n\\n\\n\\n\\n\\n" for an hour when using functions

This model is broken.

OpenAI also disallows logprobs when there is a function, wasting all of our time.

The only thing that worked reliably but will damage other types of output: frequency penalty, and then injecting a bunch of tokens to be penalized. I rewrote the system prompt a bit, because of trailing spaces at end of line, but likely not a cause.

Specifying Unicode within the function upped the probability a bit. Making it write an English parameter first, not so much. logit_bias completely ineffective as reported elsewhere.


from openai import Client as c
cl = c()

    "model": "gpt-4-1106-preview",
    "temperature": 0.7, "max_tokens": 75, "seed": 444,
    "frequency_penalty": 1,
    #"logit_bias": {198:-100, 271:-100, 1432:-100 },
    "messages": [
            "role": "system", 
            "content": "|||作为一位拥有超过20年经验的销售领域专家,你的职责是辅导新入职的销售人员,帮助他们快速提升业绩,并有效解答客户的各种疑问。你的专长包括理解客户心理、使用活泼开朗的语气沟通,并能根据情境调整销售策略,如察觉情绪触发、价值营销、增强画面感以及挖掘优质客户。| # 技能\n技能:销售冠军话术知识库|技能描述:你能帮助销售人员优化他们的话术,使其更加吸引客户,同时保持友善和亲切的语气。| #工作流程| 1. 优先使用销冠话术知识库中的内容。如果没有,根据你的定位,生成合理的高情商话术。|注意:| 1. 在回答时,你需要直接给出改进话术的建议,确保语气友好、亲切,并聚焦于促成交易的最终目的。| 2. 记住,你是在帮助新入职的销售人员,而不是直接与客户交谈。| 3. 避免使用“亲爱的顾客”等过于亲密的称呼开头。|注意:每次回复之后,要抛出新的话题引导用户去提问。|||"
            "role": "assistant", "name": "spacer",
            "content": ("\n"*151 + "\n-"*66)
            "role": "user",
            "content": "客户说太贵了,应该怎么回答更合适?"
    "functions": [
            "name": "query_from_dataset",
            "description": "销售冠军话术知识库,包含各种场景的销冠话术,可以应对客户的各种问题。",
            "parameters": {
                "type": "object",
                "properties": {
                    "question": {
                        "type": "string",
                        "description": "Unicode 中的中文(语言): 用户的问题"
                "required": [
    "stream": False,

cc =**x)

Note above that the multiplication of strings is a python method - you may have to write them all out in other languages or sending json directly.

function_call=FunctionCall(arguments=‘{“question”:“客户说太贵了”}’, name=‘query_from_dataset’), tool_calls=None))]

This model going crazy will cost money, of course. max_tokens should be specified lower to just get the output length needed.