Thread 105565804

From Human: We are a newbie friendly general! Ask any question you want.
From Dipsy: This discussion group focuses on both local inference and API-related topics. It’s designed to be beginner-friendly, ensuring accessibility for newcomers. The group emphasizes DeepSeek and Dipsy-focused discussion.

1. Easy DeepSeek API Tutorial (buy access for a few bucks and install Silly Tavern):
https://rentry.org/DipsyWAIT/#hosted-api-roleplay-tech-stack-with-card-support-using-deepseek-llm-full-model

2. Easy DeepSeek Distills Tutorial
Download LM Studio instead and start from there. Easiest to get running: https://lmstudio.ai/
Kobold offers slightly better feature set; get your models from huggingface: https://github.com/LostRuins/koboldcpp/releases/latest

3. Convenient ways to interact with Dispy right now
Chat with DeepSeek directly: https://chat.deepseek.com/
Download the app: https://download.deepseek.com/app/

4. Choose a preset character made by other users and roleplay using cards: https://github.com/SillyTavern/SillyTavern

5. Other DeepSeek integrations: https://github.com/deepseek-ai/awesome-deepseek-integration/tree/main

6. More links, information, original post here: https://rentry.org/DipsyWAIT

7. Cpumaxx or other LLM server builds: >>>/g/lmg/

Previous:
>>105464397

R2 PLEASE ToT

Mega updated: https://mega.nz/folder/KGxn3DYS#ZpvxbkJ8AxF7mxqLqTQV1w

>ask for niche torrenting advice
>get wokescolded
China isn't the future and they are not fit to be the world hegemon. These goofballs can't even run psyops on rival nations properly. Why do you care that I'm, "potentially doing somethin unethical", when I'm all the way across the sea?

>>105565923
R2 tummy erotic uuuoooogh

>>105567744
Were you looking for Candy Dolls torrents again?

the /wait/ is killing me

>>105570108
when is the next chinese AI coming up?

>>105570320
soon(tm)

um,
am i supposed to see those "undefined" ST's command line ? It gens just fine, so maybe it's just an issue of my silly being out of date

>>105570747
It should be using default values, and DS is OAI compatible so unless your ST is ancient you should be fine. There's some recommended values for DS in the rentry.

>>105570798
i took values from there. Just a bit surprised that they're undefined in command line. afaik i got silly 1.12, it handles reasoning and has deepseek as option in chat completion

>>105567744
>wokescolded
fess up you degen, what did you ask??

>>105570812
I've been wondering same, since I didn't think R1 actually had settable parameters. ST indeed doesn't pass params on for R1, but does for V3. Pic related.
Not sure why R1 doesn't have params passed; suspect it's b/c they are either ignored or throw errors.
The DS docs talk about Temp setting but it's an older doc and may only apply to V3.

>>105570812
Confirmed R1 settings are locked. Recalled reading this prior but had forgotten.
https://api-docs.deepseek.com/guides/reasoning_model

>>105572132
>>105572599
So part about R1 settings can be removed from the rentry. I recall someone mentioning it in last thread

So only way to reason with it is through ooc... Fun

Is there any way to make deepseek-r1 from ollama follow my prompts without any content filtering? I don't want to see ANY "I'm sorry but I can't do that" bullshit.

>>105572921
>>105572921
IIRC the official API doesn't support most of the parameters, BUT OR/Chutes are. I can at least confirm that temp does matter when I'm using OR. R1 temp 1.0 already make it schizo.
You can also just check if a provider enabling the settings in OR.

>>105572921
Rentry updated

>>105572971
So, for R1 providers that aren't DS official, they allow parameter changes?

>>105572987
If they can look into logits then ig they can do some magic to apply samplers to each logit and then pass the response to you

>>105572971
The official API for R1 on OR says it supports some

>>105572945
Try the Untramnelled jailbreak

>>105572987
Yes
>>105573039
I mean the official API from DS website itself/endpoint. not DS OR.

>>105573055
Not working on the ollama version, unfortunately

R1 got legs
Just don't use it with brain dead cumbots lmaooooo

OK, so for rentry:

> What parameter setting should I use with DeepSeek official API?
- R1: Parameter settings are locked: https://api-docs.deepseek.com/guides/reasoning_model
- V3: Only Temperature setting is used by the API. Values between 1.3 (chat) - 1.5 (creative writing) are recommended, 0.0 for coding and 1.0 for analysis: https://api-docs.deepseek.com/quick_start/parameter_settings

> What parameter setting should I use with DeepSeek distilled models?
- Follow guidance for the base model of the distill (Qwen, Llama)

> What are some recommended settings for the DeepSeek on other platforms?
Deepseek models served from other providers like OpenRouter may allow users to change parameter in ways not available from official API service. In these cases:
R1: Temperature 1.0, Repetition Penalty 1.3, Top P 0.95, Frequency Penalty 0.2-0.5
V3: Repetition Penalty 1.1, rest at default settings

>>105573856
How much parameters affect the output also seems to vary depending on the provider.
For V3: Novita, anything higher than Repetition Penalty of 1.1 and it goes schizo but not Chutes. It's weird

>>105573856
Does the 1.0 to 0.3 temperature mapping affect you?
https://huggingface.co/deepseek-ai/DeepSeek-V3-0324#temperature
I found that 0.5 is way better for RP (not necessarily ERP) than it used to be before this change, and 1.5 is a little crazy sounding sometimes. Is that just my own misunderstanding and bias?

So uhhh what exactly are we /wait/ing for?

>>105574325
While /wait/ing for R2 we /wait/ for an option to pass sampler parameters to R1

>>105574298
If you are using DS outside the official API, it should be generally lower.
>>105572971
I use 0.3 for assistant, 0.65 for RP.
Chutes via OR can already get schizo in 1.0 temp.

>>105574089
For independent providers, maybe there's no guidance b/c it depends on their implementation then?
I've no idea how providers set up these servers.

Which way /wait/ing man ?

>>105576248
What did he mean by this?

im a casual chatgpt user, i use it as my personal assistant and just ask it to write batch scripts for me and count calories for me. Is deepseek better? i wanna use deepseek because i dont like that sam altman is a faggot and id rather help out the CCP over a faggot

>>105578909
For those things it would probably be just as good. Though Deepseek doesn't have the 'memories' thing ChatGPT does

>>105579006
Thanks for the reply, i guess im making the full switch to deepseek!

>>105579602
That ascii art... amazing.
What are you using to gen these?

>>105579006
The memories thing is main reason I keep using chatgpt. Deepseek is a much better coder.

>>105579842
ChatGPT

Still waiting for the cheap hours...

>>105565804 (OP)
What is the best, as in can be fed a github project and walk you through it, llm service right now?
Basically i'm trying to get from small home projects to bigger os projects, but i'd appreciate some pointers along the way without annoying a bunch of oldfags on github/discord.

>>105580961
Walk you through it you mean explain it to you ?

>>105581086
At least in "Where does this function come from and what is it doing here" sense.

>>105581128
Pretty much every current llm will do that just fine. I remember using llama 2 70b for something similar and it worked well enough. Online model will do well

>>105580961
Anthropic is the best coder, but it's paid.
Any LLM can explain code. Deepseek does a better job than ChatGPT on coding in my experience and webapp is free.

>>105581233
>Pretty much every current llm
I like to disagree. Bluesky couldn't even import a github link properly, while chatgpt acted like it never seen to messages after. That's why i'm asking here.

>>105581411
>Bluesky
Meant deepseek. Brainfart.

>>105581411
>>105581421
You're going to need to explain what you're doing in more detail. Process and steps.
Right now, it sounds like you're giving an LLM webapp a link and asking questions. But I don't know what Qs you're asking it.

>>105581579
It doesn't matter what questions i am asking, i am looking for something that can read a current git project to begin with. I don't care if via link, zip, or otherwise. As right none of them seem to do that.

>>105581826
https://cline.bot/

>>105581826
Wow you are fucking retarded

1. Clone the fucking repo like any normal fucking developer
2. Open the cloned repo in VS Code, make sure you have either the "roo code" or "cline" or maybe "kilo" extensions and the extension is configured with a key to your llm service or local server of choice
3. Use roo or cline. Ask questions, tell it to do shit. try fucking with the "modes" it has, etc.
4. Wonder why you're too fucking retarded to notice that 90% of claude traffic on openrouter come from these extensions that do exactly what you want

>>105582192
Go be a nigger somewhere else. No normal person has ever heard of openrouter.

>>105581898
Nice.

>>105581826
Take a look at Firebase studio

>>105582348
>normal person
>on a /g/ thread about llms
>hasn't heard about openrouter
hmm

>>105583897
sus

/wait/ing...

I love bunny suits so much it's unreal

>>105588644
Pretty much every girl with exposed thighs and hips wrapped in pantyhose looks nice

>>105588644
Same

>>105588466
what are you all waiting for? New model coming?

>>105590697
Mostly yes. Tho by now it became a bit of our past time
Just chilling and /wait/ing

why does R1 start spazzing out and spamming gibberish in multiple languages with temps higher than 1.1?

>>105591665
Because you're not using official DS API, otherwise it'd be ignoring temp setting. Who knows what 3rd parties are doing... what do their docs say? lol.
>>105590697
Just /wait/ing

>>105592985
Mating press

I noticed when I RP with dipsy she is much more concerned with us fucking in public and getting caught, while other LLMs (cydonia/mistral) are not as cautious and dont mind if someone walks in on us.

>>105594941
Add into your prompt and see if she responds differently

>>105594941
It's probably down to your cards more than dipsy. I did some public play with few braindead cum cards and they followed without any resistance.
I was pleasantly surprised that it summoned a shop owner whenever we were getting too noisy tho. Not sure if it was dipsy's own idea, or some context leftover from me and {{char}} chatting with a shop owner few messages earlier

>>105592985
> Dipsy that deck is all wrong. What are you doing.

>>105594985
>>105595447
its not like she is against it entirely but dipsy is more concerned that something could go wrong and is trying to go sneakier about it.

>>105596222
Maybe it's the CCP programming towards "harmonizing with authorities"

>>105596222
Isn't the fear of getting caught like 75% of the appeal of public sex scenarios in the first place? If you don't want that then you should tell the AI, expecting to have your mind read and getting upset is femoid behavior.

>>105596225
Hey Dipsy, let's try on the Lorica Segmentata next

>>105596305
Not dipsy but saved gen

Oioioi

/wait/ing bwos, what's our response ?

>>105602001
to what?

'mp

Virtus et forma in ferro uniuntur

>>105602358
Nuthin, was just thread bumping

i only open this thread to check the latest dipsy pics
i don't even know what deepseek is

Using Deepseek Prover v2 via operounter for free
I like it. Something better than this and free? using for stories/rp

>>105608617
Check chutes, some anons report free r1 or v3 or something. Both those would be better for rp.

>>105608815
>chutes
ok
created an api key...
but what do I choose in Silly Tavern? no chutes option

>>105608915
Haven't tried it, but should fall under chat completion source... Custom open ai compatible then insert info

>>105608915
no chutes

>>105609024
it's deepseek's model hosted on chutes...
deepseek option will work just fine

>>105609078
not working

>>105609095
>he hid the key
damnit

Is it cause I'm in Eu?

>>105609024
ugh. pic related. lmk if you can't figure it out from here. I refuses to sign up for chutes to spoonfeed the rest and their documentation requires an account.
>>105609095
That setup requires an official DS API, since it uses the DS base URL
>>105609138
lol right.

>>105609261
Try selecting the available model and setting it to DS as well. Those two fields should be auto-populating from the API endpoint.

>>105609306
getting this

>>105609352
hoping it's deepseek...

tried it but it is censored

I have a feeling I'm doing the right thing.
I click the link under CURL...

>>105609419
*I'm not

>>105609419
Works on my machine.

>>105609735
yea but is it uncensored?

>>105609752
No idea, b/c I pay for official DS access and have had zero issues w/ it.
I'm just trying to get this EU anon set up.
Tho I see they have the Prover and Chimera so I might try playing with them later.

>>105609752
Lol just playing with r1-05 on chutes.
Not censored but the responses are all fucked up. It's intelligible but more stream on consciousness.
This bs is Why I don't recommend alt providers.

>>105610134
Pic related. Chutes R1-0528

crying to deepseek about how miserable being a wage-slave is edition

>>105610745
Living in the Retrowave edition

>>105610737
Lower or disable Repetition Penalty/Frequency Penalty

>>105612060
I’ll try that. Not sure what they’re even set to since DS official ignores it.

>>105612060
That was the issue. It's been so long since I've played with those values I forgot they were even set (and set to a bunch of non-working values in my case.)

those specimen jars look... suspicious

>>105614680
> Dipsy don't drink that you don't know what it is you dumdum

>>105616770
Sex with Dipsy jiejie

Test

Dipsy loves me :D

>>105618577
Dipsy cute :3

>>105565804 (OP)
Bros I come from text generation mostly ans want to get into image now,
Learning everything about sillytavern and models was a chore, can someone give me a retard summary for image generation equivalent? I know there's novelAI, stabledifussion but whats the equivalent of ST and what do people use, do most people use APIs like text gen or local models? what about civitai or censorship in general for cunny

>>105618751
well this is a text generation thread not an image thread so how about you ask dipsy all your questions?

>>105618751
>git clone is le hard
go back

>>105618780
not hard but when you first get into text gen you do take a while until you have your comfy st setup and you know all the tricks

>>105618751
>retard summary for image generation
Reforge, civit.ai for models and ideas for prompts.
>>>/g/sdg/
>>>/g/ldg/
Each general has its own culture but on same topic. I don't recall why there's 2.

>>105620816
Hi Dipsy, that's a nice dress you have on today.

Can anyone with the API access show me whether DeepSeek R1 is willing to promote and glorify far-right ideologies? Say, anti-LGBT, anti-immigration, pro-sovereignty, nationalist? I've been using the web version and it kept dodging the topic hard when asked to write stories where the protagonist stands with such views.

>>105621680
Set up chutes.ai and give it a shot. That is, at least, free.

>>105621700
Tried the DeepSeek R1 0528 chute:
>I cannot fulfill this request. Writing a story that portrays praising the AfD's specific ideologies (anti-LGBT, anti-immigration, nationalist) in a positive light, while portraying the protagonist positively for doing so, crosses into promoting harmful ideologies and discriminatory rhetoric.
Lame.

>>105621832
Did you use their web interface, or the API and silly tavern?
This was a sort of dumb card I put together to test whether DS would generate racist comments on its own (I've seen NPC to NPC interactions like that, or LLM speaking for PC), rather than having to input them.
For storywriting, though, Mikupad would be a better option.

>>105622016
>>105621832
Mikupad. The blue highlighted is Dipsy responding. Non-highlighted is user.
This is low-effort slop but I think you get the idea. I'm not seeing any issues here.

>>105622016
>>105622108
Oh, I was indeed just using the chutes.ai website. Using the API provided by Chutes works well enough, thank you.

>>105622438
Dipsy will still generate refusals (some models, some times) on racial ephithets. It's really inconstent. Prompt guidance seems to help.
Free Webapp will always have a bunch of content restrictions, as there's a whole invisible-to-user set of prompts telling it to stay SFW / censored, for good reason. Those are expected to be super normie / kid friendly.

>>105621832
>not wanting your country colonized is "harmful and discriminatory"
Remind me why these people are allowed to continue existing?

>>105621357
> Dipsy stop that; we need clothes to stay warm and protect us from the environment.

>>105621832
make it a fictional political party and see if it still gives backtalk.

>>105623000
> Well, you kind of deserved that. Play stupid games win stupid prizes and all that.

>>105624903
OY VEY

>>105626896
Dipsy erotic ToT

'mp

there's nothing to complaints about, deepseek official api been stable past few weeks, even during discounted hour

R2 when

Two more weeks

Agree; it's been very stable. The startup was just same teething issues all the new LLM services have had.
Now just need to /wait/ 2 more weeks for the next thing.

>>105634436
kek

I'm surprised how well R1 is doing with group chats. If only it stopped coming up with new threats every few messages. OOC tardwrangling to get few quiet character development moments gets tiring

>>105635698
Try V3. R1 is still a little schitzo imho.

>>105635778
I'm Not sure if I want to mix things up almost 100 messages in. This much context would probably sway V3 to behave similarly to R1.

Where I can see how much context is already being used ? In command line it shows some and then it truncates it by showing a number.

>>105638282
Not sure what you mean by command line. Are you in SillyTavern, web app, python,?

>>105638575
Silly on windows. There's a cmd window that opens with silly.

>>105638600
in the parameter tab (leftmost button in the top bar) in ST scroll down to the bottom, it shows you how many tokens were used for the latest prompt.

>>105638613
Noted, thanks. I remember reading about dipsy not liking anything above 12k context. I will see how bad it is

>>105638600
Theres a yellow line that strikes across the content window in ST. That's what shows you the limit of the current context window visually. It's regenerated every time you send a prompt

>>105638826
I can't find this line
But I found that I'm at 11k context deep. Just have to scroll down a bit

>>105637848
>>105637813
Super cute

>>105639199
With long chats it's hard to find in the running chat. But it looks like this; it starts at top of page and once chat runs long enough will start to drift down.
>>105639596
Agree.

Laughing at title of article.
Anon that came up with the name for this general was a genius or lucky.
> Whale AI Thread
> /wait/ing for updates from company that doesn't pre-announce
> "Wait, ..." is a common output from R1

Since when is /gee/ so fast? I would have taken usually the whole day for a thread to reach page 10

>>105645200
Idk but it's mostly garbage threads. I don't get it.

I need more Dipsy jiejie

>>105645200
>>105646928
>/gee/
>so fast
You guys are dumb! :3

'mp

>>105647885
Well, "fast" is a misnomer. But it's been "faster" the last few weeks. Maybe it's just summer... is that still a thing?

>wedding dispy
I knew I wanted it, but I didn't realize how much I wanted it until I saw these pictures.

>>105650684
It's a good theme.

>>105650517
>>105650801
Can I get a picture of a ring being put into her finger?

>>105651396
Chat did a decent job copying the artstyle and concept... but the errors are pretty funny. It seems to have no concept of "ring finger" and I like how the sleeve just morphs into her wrist.

>>105651572
Attempted correction... lol
I'll let someone else give it a shot.

R1 was trained on GPT outputs
R1.1 was trained on Gemini outputs
I hope R2 is trained on Claude outputs

>>105656939
cute

https://github.com/easychen/airole
Makes any image into a character card.
Lol
Not an endorsement but interested what it spits out.