Would anyone be so kind as to recommend models?

[Chimal1994]

Would anyone be so kind as to recommend models?

I've been using the ones in the workshop for a while (Gemma 2 and some image generation ones whose name I don't remember and it's not very relevant XD) and I also ended up trying Gemma 3, but it was disappointing. It ignored things and didn't track some variables. Gemma 2 turned out to be better, ironically.

In the workshop, I see stories with really nice images, and I'd like to know what models they use. Also, if you know of a text model that generates better responses than Gemma 2, I'd appreciate it.

I also want to clarify that I use a toaster with a 9th gen Intel Core i5 and a GTX 1660, so if you can recommend something with that in mind, and if not, then that's fine too. XD Let's try it, it's been said.

[Oleg Skutte]

Originally posted by Chimal1994:

By the way, would it be possible to expand the character limit of the status box (the one where we put which variables we want to track) in a future update?

Character limit of the status box is tied to context window size. If you're already at max context size you can edit the value with regedit.exe to get past the limit for now. I might increase the upper limit for context window size of local inference in the future.

[Topper Harley (toppertrip)]

Originally posted by Chimal1994:

Originally posted by Topper Harley (toppertrip):

you know, a potato can power a powerful AI, GLaDOS

GLaDOS owes me a cake!

Note that you can perfectly ignore: I rewrite this because I don't want to fill this discussion with meaningless comments (I hope I don't send you a notification for rewriting something where I mention you and if yes, then sorry Topper)

Thanks to Portal's references I found that there is a model to make GLaDOS narrate the story XD

His voice makes your head hurt but for a Portal fan this is gold! You play DREAMIO and GLaDOS tells your adventure xD

GLaDOS TTS is called

Interesting if you can find a way to use local models properly. Another user wrote here about changing the seed to get different results for the same user responses

[Chimal1994]

Good morning, lovely people of the Dreamio forum. I've been testing out models and found one that's quite interesting and promising.

It's small and fast, gives pretty good responses, and, as the cherry on top, it's uncensored (important for those adventures where you don't want the narrator constantly reminding you that you're doing something immoral), but it does have a catch... The context window is small, about 2k (for me, specifically, since I like to put a lot of information like weather, time, day, and year in the status box, it's a limitation, but I know for many others who don't have such a cluttered status box, this won't be a problem).

The model is called Dolphin 2.6 Phi-2.

I think the model is perfect for all of us who don't have a NASA PC, you know, for those rainy days or when you're traveling to a place without internet and you want a story on Dreamio.

[Chimal1994]

Interesting if you can find a way to use local models properly. Another user wrote here about changing the seed to get different results for the same user responses [/quote]

Forgive my ignorance, but I'm not entirely clear on what you mean by seeds.

I've noticed, for example, that if you activate the reasoning option in Dreamio, the model will give you varied and different responses, even if you tell it the same thing. For example, if you say "Hello, how are you?" and it responds "Hello, I'm fine" or "Good morning, I'm great."

Although, of course, the same option warns you that depending on the model you use, it may work better or worse.

I've also noticed that versions of the same model respond differently, as is the case with Gemma 2 and Gemma 3 (if you're wondering how I noticed this, it's because I always start the adventure by specifying who I am and what I look like).

And obviously, different models speak differently (Captain obvious, reaffirming the obvious).

It also greatly affects the instructions you give the narrator. For example, one season I wanted to make an immortal protagonist, but the default narrator always killed me anyway, even though according to the story, I shouldn't be able to die under any circumstances. I managed to make my character immortal by modifying the narrator's instructions. If you feel like the narrator is repeating himself too much, you can always Add a couple of instructions indicating that your answers should be more creative.

And I almost forgot, if we have a very large context window, the narrator tends to repeat himself and pay less attention to us (although I don't mind and I have it at 4096 xDD, Dreamio recommends having it at 2048 by the way)

[Topper Harley (toppertrip)]

There is a number associated to the generation process,

here what Chatgp says about "seeds":

Originally posted by ChatGpt:

In the context of artificial intelligence, particularly in language and image models, a "seed" refers to an initial number used to generate random sequences. This number plays a crucial role in the sampling process for generating responses or outputs from a model.

Text Generation: When generating text, the seed determines the starting point for a sequence of pseudo-random numbers that the model uses to decide which word or phrase to generate next. By changing the seed, you alter the sequence of numbers produced, leading to different outputs even if the input remains the same.

For instance, if you input "Hello, how are you?" and the model is set to a specific seed, it might respond with "Hello, I'm fine." However, if you change the seed, the model might generate "Good morning, I'm great" for the same input. This variability is essential for creative applications, allowing for a range of responses that can make interactions more engaging.

The seed can be set in the model's parameters before generating text. In many programming environments or APIs, you can specify the seed value as an argument when calling the function that generates the text. This allows for reproducibility; using the same seed and input will yield the same output, which is useful for testing and debugging.

Image Generation: In image generation models, the seed functions similarly. It initializes the random number generator used to create the image. When you input a prompt for an image, the seed influences the various aspects of the image creation process, such as colors, shapes, and overall composition.

Changing the seed in image generation can lead to significantly different visual outputs from the same prompt. For example, if you ask for "a sunset over a mountain," the model might produce vastly different images based on the seed value, such as varying colors or styles of mountains and skies.

Just like in text generation, the seed can usually be specified in the model's settings or API call. This allows users to explore different creative outputs or maintain consistency in their results when needed.

Originally posted by Chimal1994:

I've noticed, for example, that if you activate the reasoning option in Dreamio, the model will give you varied and different responses, even if you tell it the same thing. For example, if you say "Hello, how are you?" and it responds "Hello, I'm fine" or "Good morning, I'm great."

And here chatgpt about "deep reasoning", it says there is a random factor inside the reasoning process:

Originally posted by ChatGpt:

When you activate the reasoning option, the model is likely employing a more complex decision-making process to generate responses. This process can involve:

Dynamic Contextual Understanding: The model analyzes the input not just for its literal meaning but also for context, tone, and potential implications. This allows it to generate responses that are not only varied but also contextually appropriate.

Stochastic Sampling: The model utilizes stochastic sampling methods, where randomness is introduced in the selection of the next word or phrase. This is often controlled by parameters such as temperature and top-k sampling:

Temperature: A higher temperature setting increases randomness, allowing for more diverse and creative outputs. Conversely, a lower temperature results in more deterministic and predictable responses.
Top-k Sampling: This method limits the selection to the top 'k' most likely next words based on the model's predictions. By varying 'k', the model can either focus on the most probable responses or explore a wider range of options.
Seed Variation: As previously mentioned, the seed plays a crucial role in generating responses. If the seed is set to a specific value, the model will produce consistent outputs for the same input. However, if the reasoning option incorporates dynamic seed adjustments or if the model uses a different seed for each session or response, this can lead to varied outputs even for identical prompts.

I used it with image generation and I got different kind of images. For example with low numbers I got a "flat" image, almost 2D, with higher numbers I got something similar to 3D, the version I used suggested to insert higher number for a higher image quality generation.

[Topper Harley (toppertrip)]

Btw, I use the term seed, but shouldn't be the right one, here is the topic about the Ai temperature:

https://steamcommunity.com/app/2795060/discussions/0/596269063005585099/

[Chimal1994]

Originally posted by Topper Harley (toppertrip):

Btw, I use the term seed, but shouldn't be the right one, here is the topic about the Ai temperature:

https://steamcommunity.com/app/2795060/discussions/0/596269063005585099/

The truth is that now that you explained the issue of temperature in AIs, it wouldn't be bad to have it, but it depends on Oleg.

If I understood correctly, we do not even have the option if we use local models xD

Oleg please add it

[Oleg Skutte]

Temperature in Dreamio is locked at 0 because increasing the temperature introduces randomness in model responses, which often inhibits the ability to follow instructions and maintain required formatting, increases hallucinations, and generally negatively impacts reliability. Essentially, higher temperature puts sticks in the model's wheels.

Temperature is set higher for summary generation and status generation only, in order to combat repetition.

[Topper Harley (toppertrip)]

Thanks Oleg, I was wondering how temperature works.
In another topic someone wrote about using ComfyUI to set temperature for local models.
I knew it was used mainly for image generation.

[Chimal1994]

Originally posted by Oleg Skutte:

Temperature in Dreamio is locked at 0 because increasing the temperature introduces randomness in model responses, which often inhibits the ability to follow instructions and maintain required formatting, increases hallucinations, and generally negatively impacts reliability. Essentially, higher temperature puts sticks in the model's wheels.

Even so, I think it is a good idea to place the option and let whoever wants to use it.

You can always have it disabled by default with a warning, surely more than one will enjoy modifying the temperature.

Even if it is in a very, very distant update, I would like the option to eventually appear.

[Oleg Skutte]

Dreamio actually had a temperature setting when it first came out. I removed it after getting negative reviews from people who had bad experiences after increasing the temperature.

[Chimal1994]

Originally posted by Oleg Skutte:

Dreamio actually had a temperature setting when it first came out. I removed it after getting negative reviews from people who had bad experiences after increasing the temperature.

Wouldn't placing the typical Dreamio message recommending that you have the temperature value in (x) solve the problem of people modifying it and not knowing how to return it to how it was?

You already have one of those pop-up messages in the context window configuration

[Chimal1994]

I have come up with the perfect solution for the temperature to return, which is in a section called advanced within the text generation settings and that's it, most people never touch anything with the name advanced in the settings menus xD

[Oleg Skutte]

Temperature setting has been added in the latest update.

[Chimal1994]

Good morning people from the Dreamio forum.

I started testing the models again and noticed something curious Gemma 3 12B continues to give strange responses, but oh god Gemma 3 27B gives excellent results, although sometimes it stays in a strange loop tokenizing to infinity but I guess it's more than anything to run a model of that caliber in my toaster, I also tried Gemma 3 1B and 4B but as was obvious they are not a good option, they are fast but their responses do not make sense and they quickly forget the plot

For now, for me, things look like this:

1.- Gemma 3 27B
2.- Gemma 2 27B
3.- Gemma 2 9B
4.- Gemma 3 12B

I still don't understand why Gemma 3 12B is worse than Gemma 2 9B :/

By the way, thanks Oleg for the temperature, it's really fun to increase it and see how the answers become more creative or even see that the AI ​​reveals itself and founds Skynet