Stop sequence in the model does not correspond to what I sent?

My model:

{“prompt”: “Come posso trattare la dispensa e i pensili?\n\n\n”, “completion”: “Spruzzando una giusta quantità di prodotto, spostando accuratamente gli alimenti e riponendoli una volta effettuato il trattamento.###”}

Come posso trattare la dispensa e i pensili?

			$$$ Spruzzando un po' di prodotto su uno spugnolo, lo si applica sulle superfici interessate.


			$$$ Riporre alcuni prodotti nella dispensa e altri nei pensili, per evitare che gli insetti li contaminino.


Are you using chatgpt and have a bug?
Are you using the API?
Are you attempting to fine-tune a model?

If you are fine-tuning the new models davinci-002 or babbage-002, via the new endpoint for fine-tuning, we finally have an example of a training set for those models:

{"prompt": "<prompt text>", "completion": "<ideal generated text>"}
{"prompt": "<prompt text>", "completion": "<ideal generated text>"}
{"prompt": "<prompt text>", "completion": "<ideal generated text>"}

There is no additional text.

If you already tuned on that input, you’ll likely have to add three \n\n\n to all your user questions. And watch for and strip the ### if you receive them. The training may have been too damaged still.

Training the base models on how to be a chatbot without a input format that makes the completion model “complete” will take a LOT of examples. A minimum prompt (but write with \n):

A chatbot answers questions:
User: how can I train an AI?

I use \n\n\n to end the prompt and ### to end completion, why do I see $$$ in the response?
“choices”: [
“text”: “\n\n$$$ Spruzzando una giusta quantità di prodotto, spostando accuratamente la microfibra sulle superfici, senza strofinarle.$$$\n\n$$$$$$\n\n$$$$$$\n\n$$$$$$\n\n$$$$$$\n\n$$$$$$\n\n$$$$$$\n\n$$$$$$\n\n$$$$$$\n\n$$$$$$\n\n$$$$$$\n\n$$$$$$\n\n$$$$$$\n\n$$$$$$\n\n$$$$$$\n\n$$$$$$\n\n$$$$$$\n\n$$$$$$\n”,
“index”: 0,
“logprobs”: null,
“finish_reason”: “length”

I have a suspicion that when using the “prompt”/“completion” format, that OpenAI are injecting the prior manually-created stop sequence themselves, and have chosen the $$$.

The model may be completely confused by being trained on the ### and the carriage returns.

If you trained gpt-3.5-turbo on the completion format instead of the role messages, that also could explain the nonsense.

You can increase the repetition and presence penalty, or better, identify the b100k token numbers that are being output and demote them with logit_bias, but I think safe to conclude that the model will not be useful.