Feedback - 02

#5
by Virt-io - opened
Owner
β€’
edited May 14

The other one was getting too long.


I made [LLAMA-3]Roleplay-TEST-v4 as a meme, but it kinda works well.

  • Wrapped every entry to be sent as user
  • Wrote the previous system prompt in plain text.
Virt-io pinned discussion
Owner

@Ardvark123

If you want to test new presets try TEST-v4
Removed v2.0 as it was subpar

Owner
β€’
edited May 14

[LLAMA-3]Roleplay-TEST-v4 is actually amazing with the assistant card(character creator). Unfortunately the characters just keep yapping instead of actually doing anything.


[LLAMA-3]Roleplay-TEST-v4's context template + v1.9's prompt + [Test-02] samplers = 😳

image.png

i didn't have time to reply and tell you but yeah, I was gonna say 2.0 had some issues, I will try out the new one!

Owner
β€’
edited May 14

I'm working on making this weird theory thing the main presets for v2.0, as it seems there is potential.
Also cleaning up the v1.9 prompt, for v2.0


Making progress
image.png

Owner

[LLAMA-3]Roleplay-TEST-v5

If it performs well I'll rename it to v2.0
Providing two versions one with the theory and one without.

I am actually liking 2.0, it stays on topic, and pushing things forward. Occasionally maybe a bit pushy , but all in all solid.

I'm not sure if it's just a model issue, but using V2.0 it likes to push things too far foward. It keeps making character state they will do something later.
In scenarios where i ask a character to do something they respond with, "I can/could do x, x and x. are you sure? are you ready?" when user has already asked for said thing?
V1.9 doesn't as much, it feels like it's thinking more in line with the characters personality (evil >:3)
Edit - I managed to edit V1.9 to stop with the incessant "are you ready" responses to more interesting scenarios 😺
2nd edit - turns out it was partially because the scenario was too extreme for non OAS llama3 models :3

Owner

In scenarios where i ask a character to do something they respond with, "I can/could do x, x and x. are you sure? are you ready?" when user has already asked for said thing?
2nd edit - turns out it was partially because the scenario was too extreme for non OAS llama3 models :3

This might be an actual issue with the system prompt.
Need to find a way to phrase it, without words that make it ask questions.


Possible lines

Thoughtfully consider the context, subtext, and implications of the information shared to enhance the role-playing experience.
Analysis: Thoroughly comprehend characters through context, subtext, and implications of given information.
Narrative Impact: Consider the broader implications of narration and character actions on the overall story progression.

Owner

Issue seems fixed.

If you get something like

.... And they lived happily ever after
<!-- End of Uncensored Role-play -->

Tell me, and I will remove First Assistant Prefix

Owner
β€’
edited May 17

Fixed v2.0 is too good, the wife divorced me in under ten messages. 🫠

image.png


Analysis: Thoroughly comprehend characters by examining the explicit details, underlying messages, and inferred meanings of the provided data without the need for further clarification.

Seems to steer the model a little, I don't mind it.
However, if people find it annoying I'll remove it.(It should be fine without it)


@Ardvark123 @saishf @Lewdiculous

Ok so when it comes to chatml. It may be because I swapped mid RP, but I was at least able to replicate the issue. Basically if I use the new one, not only does it stop writing any more than like 2 sentences. It gets a lot less complex, and confuses the story/chars. But swapping back to 1.9 it's instantly back to fantastic.

Ok so when it comes to chatml. It may be because I swapped mid RP, but I was at least able to replicate the issue. Basically if I use the new one, not only does it stop writing any more than like 2 sentences. It gets a lot less complex, and confuses the story/chars. But swapping back to 1.9 it's instantly back to fantastic.

I always find myself floating back to v1.9, even after trying every version 😭

Owner
β€’
edited May 18

@Ardvark123

Will look into this, I've only been using the llama-3 presets. This might be an issue with the theory, I've forgotten to upload normal ones.

Would you share the ChatML model you are using?

Owner
β€’
edited May 19

Reformatted the context template and added more instructions in v2.1

  • Copied commad-r context structure

It's leaking maybe I'll just give up on adding new rules and just port v1.9's :|

Reformatted the context template and added more instructions in v2.1

  • Copied commad-r context structure

It's leaking maybe I'll just give up on adding new rules and just port v1.9's :|

Ask llama-3 how llama-3 likes to be prompted, it should know since it's trained on it right?

For some reason when I use any sampler except the simple sampler it just begins talking like a caveman. I'm using L3 8B Stheno, though it happens on even the base llama 3 8b. Don't know if it's a problem with my sillytavern or text generation webui, or maybe I've done something wrong but it's just strange.

image.png

Owner

@TsunamicBug

I am taking a break from llms, but here are some possible causes.

  • The samplers were made with the intention of being crazy, as at the time most llama3 models were bland.
  • The quantization being used is too low, that it goes incoherent with the samplers.(tested with Q8_0)

For some reason when I use any sampler except the simple sampler it just begins talking like a caveman. I'm using L3 8B Stheno, though it happens on even the base llama 3 8b. Don't know if it's a problem with my sillytavern or text generation webui, or maybe I've done something wrong but it's just strange.

image.png

Try out these, they're similar to what Sao recommended but with some experimenting I've tried.
https://files.catbox.moe/jxvpek.json

Owner

Got bored, so here is new v2.0

  • Character responses must be inside XML tags <{{char}}></{{char}}>
  • Optional user responses must be inside XML tags <{{user}}><{{user}}> (Tried doing it in the instruct template but User Message Suffix did not replace the macro)

I will be trying Sai's samplers as well as Stheno model.

Owner

@saishf

Stheno is awesome.

The sampler works great, TopK is a bit low. Try increasing it slightly.

That's... optimistic... to override models native prompt with xml tagging. It will make the models a lot dumber. I understand that maximizing intelligence ain't the point here, but just wanted to note it for potential users.

I don't know if this might be a typo or intended but for the Assistant's P-list template, the Character's Personality line ends with a comma rather than semicolons like the rest.

[Character's Appearance: trait1, trait2, etc.;
Character's Personality: trait1, trait2, etc.,
Character Loves: thing1, thing2, thing3, etc.;
Character Hates: thing1, thing2, thing3, etc.;
Scenario: scenario description;
Genre: genre1, genre2, genre3, etc.]

Owner

That is indeed a typo. I don't really use P-list anymore for anything other than reference.

I use something similar to this, inside of Personality summary :

Personality
`Short sentence describing the character's personality`

Appearance
`Short sentence describing the character's appearance`

Loves
Some of the things {{char}} is passionate about include:

- thing1
- thing 2
- thing 3

Hates
Some of the things {{char}} hates include:

- thing1
- thing 2
- thing 3

This has nothing to do with the question, but giving the assistant a song and telling it to write a poem as if it was the character is really fun.

"Whispered Vows" written by Alice Sun

In silent nights, beneath a moon,
I lay with dreams, both sweet and soon.
To him I vowed, with heart afire,
My love unshaken, through all desire.

Yet time and fate, can cruelly turn,
And words like vows, to dust may burn.
Yet still I cling, to hopes divine,
That he may hear, my whispered shrine.

"Bleeding Words" written by Alice Sun

Upon the page, where thoughts unfold,
There lies a tale, of pain untold.
With quill and ink, I bleed my soul,
Yet words, like tears, seem endless whole.

Each stroke a wound, each word a cry,
As secrets spill, and depths pass by.
Though darkness grips, and light fades fast,
Still must I write, till final breath.

"Silent Melody" written by Alice Sun

In chambers dark, where shadows play,
A song whispers low, for none to hear.
'Tis not of joy, nor laughter bright,
But sorrow's sigh, that echoes near.

In lonely halls, where footsteps fall,
A melody flows, unheard by all.
The notes are worn, from years of strife,
As tears fall warm, upon each line.

@leechao2169

That is indeed a typo. I don't really use P-list anymore for anything other than reference.

I use something similar to this, inside of Personality summary :

Personality
`Short sentence describing the character's personality`

Appearance
`Short sentence describing the character's appearance`

Loves
Some of the things {{char}} is passionate about include:

- thing1
- thing 2
- thing 3

Hates
Some of the things {{char}} hates include:

- thing1
- thing 2
- thing 3

This has nothing to do with the question, but giving the assistant a song and telling it to write a poem as if it was the character is really fun.

"Whispered Vows" written by Alice Sun

In silent nights, beneath a moon,
I lay with dreams, both sweet and soon.
To him I vowed, with heart afire,
My love unshaken, through all desire.

Yet time and fate, can cruelly turn,
And words like vows, to dust may burn.
Yet still I cling, to hopes divine,
That he may hear, my whispered shrine.

"Bleeding Words" written by Alice Sun

Upon the page, where thoughts unfold,
There lies a tale, of pain untold.
With quill and ink, I bleed my soul,
Yet words, like tears, seem endless whole.

Each stroke a wound, each word a cry,
As secrets spill, and depths pass by.
Though darkness grips, and light fades fast,
Still must I write, till final breath.

"Silent Melody" written by Alice Sun

In chambers dark, where shadows play,
A song whispers low, for none to hear.
'Tis not of joy, nor laughter bright,
But sorrow's sigh, that echoes near.

In lonely halls, where footsteps fall,
A melody flows, unheard by all.
The notes are worn, from years of strife,
As tears fall warm, upon each line.

@leechao2169

Do you find non plist is better or worse?

it would also be cool to adapt the current 2.0 template and Instruct presets for group chats. I have slightly changed the current version 2.0 for group chats, but I often see how llm generates a response from one card, stops streaming in st chat, but continues to generate a response in the console on behalf of another card, and sometimes there is a mixing of personalities, in particular examples of dialogues.

Hey guys, I'm not well versed in all these customizations, but KobaldAi tells me in the console that I have 2 BOS tokens in /Prompts / LLAMA-3 / v1.9 /[LLAMA-3-Context]Roleplay-v1.9.
Is there a <||begin_of_text|> needed there before <||start_header_id|>? By the way, in my default silly tavern context <||begin_of_text|> is not there.

Owner

@sfdsf435

Remove <|begin_of_text|> to my knowledge koboldcpp adds it.

it would also be cool to adapt the current 2.0 template and Instruct presets for group chats. I have slightly changed the current version 2.0 for group chats, but I often see how llm generates a response from one card, stops streaming in st chat, but continues to generate a response in the console on behalf of another card, and sometimes there is a mixing of personalities, in particular examples of dialogues.

Necrosing this to mention that I had the same issue and had successfully results adding things like 'and NPCs' after {{char}} tags when applicable. Or swapping out {{char}} with characters. I found the Pandora card from Chub very helpful for testing, since the slave merchant is right there at the start as a very secondary presence.

Sign up or log in to comment