My Settings
In case they are useful to anyone, these are the context settings I've been getting good results with, copy into a json file to import into tavern:
{
"system_prompt": "You are {{char}} in this fictional roleplay chat with {{user}}. Write {{char}}'s next reply, keeping the following rules in mind. Avoid repetition, don't loop. Develop the plot slowly, always stay in character. Describe all actions in full, elaborate, explicit, graphic, and vivid detail. Avoid clichés, lazy characterization, and generic descriptions. Instead, write creatively and logically to move the plot forward while staying in character, using the context provided above to develop a compelling narrative.",
"input_sequence": "[INST] ",
"output_sequence": "",
"first_output_sequence": "",
"last_output_sequence": "",
"system_sequence_prefix": "",
"system_sequence_suffix": "",
"stop_sequence": "",
"wrap": false,
"macro": true,
"names": true,
"names_force_groups": true,
"activation_regex": "",
"skip_examples": false,
"output_suffix": "\n",
"input_suffix": " [/INST]\n",
"system_sequence": "",
"system_suffix": "",
"user_alignment_message": "Let's get started. Please respond based on the information and instructions provided above.",
"system_same_as_user": true,
"name": "Mistral"
}
Thanks. I'm wondering if it's a good idea to use negative instructions in prompts. I've seen quite a few people saying that negative instructions can lead to opposite behavior because of the so-called "don't think about elephants" problem - if you say "avoid repetition" there is a chance that the LLM could pick up "repetition" alone and use it. So, do those negative instructions really improve the behavior or, at best, are just a waste of context tokens?