Excuse my retardation, but who is LEAD and who is ASSOCIATE?
I haven't had much experience with instruct models, so please, bear with me for a minute. With classical llama, the format is as follows (usually):
Persona: <>
Scenario: <>
(start token)
(example dialogue)
(greeting)
Say, I want to play out the following scenario: There is John Doe, and the player (lets call him Anon) meets him on the street and asks him for a cig. John Doe is just a regular person with no distinguishing characteristics. Usually, it would be done like this:
John Doe's Persona: John Doe is just a regular person with no distinguishing characteristics.
Scenario: There is John Doe, and the player (lets call him Anon) meets him on the street and asks him for a cig.
// (start token)
Anon: Hey man, can I have a cig?
John Doe: Sure, here you go.
(and then, the greeting, whatever that may be)
So how do I do the same thing here? As I said, excuse me for my retardation here.
Just need an example of a formatted initial prompt here
I don't have any issues using characters like this in the 13b
The source material (bluemoon) is mostly third person roleplay, so the roles LEAD and ASSOCIATE reflect that. In those conversations, one vividly narrates the actions, thoughts etc. of the character rather than playing in the first person. Used like this, the model performs its best. It can certainly do direct 1st person chatter as well, but for the best results it might be important, at least for the AI player, to have a go at this from the third person.
Using the Anon and John Doe as an example, the prompt might look like this (obviously there are much better ways to set it up, but this is just something basic):
A transcript of a roleplay between two players, LEAD and ASSOCIATE. LEAD sets up a scenario and the characters, from which ASSOCIATE then assumes a character role and continues the story for that role in response to description given by LEAD. The story and characters are developed by exchange of detailed event descriptions and character dialogs, successively given by both LEAD and ASSOCIATE.
LEAD: Hello, let's do a role play. Me as Anon, you as John Doe.
ASSOCIATE: Hello! Please describe the setting. I will pick up from there as John Doe.
LEAD: Anon, [your character description, if any], meets John Doe on the street and asks him for a cig. "Hey man, can I have a cig?", Anon asks Joe.
ASSOCIATE: John Doe is just a regular person with no distinguishing characteristics. [add here everything else describing John Doe, as well as his response in "quotes". Additionally, describe everything else he thinks of the situation or how he reacts to it. This will set up the characteristics of the character.]
LEAD: [Prompt ends here and interactive part starts: from now on, you can switch to 1st person for your character if you wish, or continue third person RP.]
In those conversations, one vividly narrates the actions, thoughts etc. of the character rather than playing in the first person. Used like this, the model performs its best.
This is definitely the trick.
I think a lot of actual prompting is unnecessary. The character + scenario setup that text-generation-webui does works pretty well. It helps a lot to include a good initial greeting or example dialog too. But I think the third person thing is really necessary, including the quotes around spoken word. This is really great in practice because the characters convey their feelings, actions, etc. It can be very expressive.
Edit: Also, as a bonus include some good roleplay or fandom elements in the scenario.