Hey all! I am writing about 200 synthetic day-in-the-life journal entries. Ideally, these are from different “people” and focus on different subjects.
Logically, when you run a generic prompt to “write a journal entry about your day, include specifics and details…” the entries come out pretty similar.
I generated a python list of 200 different jobs, ages, martial statuses, etc. Then I ran my prompt in a while loop and appended to the end of each prompt “Write as if you are a {insert specific life situation from python list}.” This effectively gave me 200 different prompts.
However, while this approach improved my outputs a little, but there is still a very generic feel and structure to all of them. Like 20% of them start with some variation of “Man, today was a rollercoaster!”
I’m wondering if people have tried to achieve something similar and what they have done?
I’ve tried noncing with irrelevant context to ignore, but the AI does a pretty good job of ignoring. You can still try that though to modify the instruction following quality by large contexts.
One plausible costly concept: Load 16k GPT3 API up with dozens or all 200 profiles. Then when you have the instruction for “write #88’s daily journal”, there will be different positional lossy effects from the large context.
Roleplay characters are all now very canned and hammy. It’s almost like OpenAI decided to add “Ah, I see” fakery detection phrases to everything that’s not a chatbot - and I get back again exactly what I expect again:
Act as a learned professor in the art of juggling mousetraps. Ah, welcome, my esteemed audience,
Curses, feeble-minded bot.
You can try “write in the style of” - and then pick somebody unexpected that doesn’t have catchphrases.
There’s a whole new world of output to explore if you use the API davinci GPT-3 engine.
You probably just need to add a “style” list and turn top_p to 1. Sounds like you have life content, but not personality content. Like “Cheery southerner”, “Serious educated business man.”
Be careful with race and region though, I’ve noticed it sometimes writes racist stuff. Specifically with asians.
Hi! You can start with something like “How do you imagine a chef in a Italian restaurant to speak?” and store it with the role. Next time you request the journal entry pass it as context and prompt the model to use the style.
Of course you can’t ask the model to imagine anything but you get the idea.
On top of using MBTI, you may be able to apply Plutchiks Wheel of Emotions to have the bot choose an emotion for the output. A sad ENTJ, a blissful ENTP, an angry ISTP, etc… we might apply to a bot we’re building, but instead of MBTI using Enneagram with wings and levels of integration… I proposed it, but we’re still work on some base functionality.
So I’ve put in a couple days on this now with some troubleshooting. I found that the models don’t understand a simple MBTI anagram very well. You get weird stuff like, “As an ENTJ I enjoy challenges in the workplace…”
What I ended up doing was getting GPT-4 to create an array of 100 unique emotional states from as general as “Angry” to as specific as “yearning for unrequited love”. Then I pick a random one each time and insert it into the prompt in a line like "The overwhelming emotion you are feeling at this moment is: {random emotion}.
I just noticed you mentioning anagram manipulation issues, this is often down to the way the model tokenizes word parts, to you ENTJ looks like 4 letters, but to the model it looks like this 2 tokens.
As a test, can you try inserting a separator character such as - (dash) to give this E-N-T-J, and the model will now see that as 4 separate tokens.
You can then extract the - to rebuild your string. See how that works for you.
I had an idea, where you just take one from each column to make your personality. This is the bot’s idea of meeting my specs.
AI: Below is a table with four characteristic categories as column headers, and 30 rows with writing examples and corresponding personality characteristics. The columns group similar facets of information together, and the entries within each row do not conflict with each other. Please note that these are fictional characters, and the provided personality characteristics are purely for creating distinct writing styles:
Personality Trait
Age & Ability
Positive
Negative
1. Optimistic
35 years old, Novelist
Always sees the bright side
Tends to be naive
2. Cynical
42 years old, Journalist
Sharp wit and satire
Pessimistic worldview
3. Enthusiastic
28 years old, Blogger
Exclamation marks aplenty
Sometimes overly verbose
4. Analytical
45 years old, Researcher
Logical reasoning and evidence-based writing
Lacks emotional depth
5. Whimsical
19 years old, Student
Imaginative and creative language
Difficulty staying focused
6. Precise
50 years old, Editor
Meticulous attention to detail
Can be overly pedantic
7. Laid-back
33 years old, Traveler
Informal, colloquial expressions
May come across as lazy
8. Authoritative
60 years old, Historian
Firm and confident tone
Tends to be domineering
9. Empathetic
27 years old, Therapist
Understanding and compassionate
Overuses qualifiers (e.g., maybe)
10. Humorous
38 years old, Comedian
Witty remarks and humorous anecdotes
May rely too much on jokes
11. Pensive
31 years old, Philosopher
Thought-provoking and contemplative
Can be overly abstract
12. Confident
48 years old, CEO
Assertive language
Comes across as arrogant
13. Nurturing
29 years old, Parent
Warm and caring tone
May be too sentimental
14. Stoic
55 years old, Veteran
Dispassionate and stoic writing
Lack of emotional expression
15. Playful
22 years old, Gamer
Uses gaming slang and metaphors
May lack depth in serious topics
16. Persuasive
37 years old, Salesperson
Convincing arguments and rhetoric
May oversell or manipulate
17. Curious
26 years old, Explorer
Inquisitive and eager to learn
Easily distracted by new ideas
18. Diplomatic
43 years old, Diplomat
Tactful and diplomatic language
May avoid taking clear stances
19. Rebellious
20 years old, Rebel
Defiant and anti-establishment
Can be confrontational
20. Melancholic
41 years old, Poet
Reflective and emotional
May lean towards melodrama
AI: You can now combine one characteristic from each column to create a unique personality for a character. For example, “joe’s writing is: Rebellious, 20 years old, Defiant, and tends to be confrontational.” This combination will create a distinctive writing style for Joe based on the specified attributes.
Obviously cripplebot doesn’t write at the length we want, but you can probably just “continue” this list, or drop a refined table header for a different list.
2 tokens.
4 separate tokens.