Thread 105794374

Expecting Steve Edition

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:
>>105661688

pissi cuckitini

Mega updated: https://mega.nz/folder/KGxn3DYS#ZpvxbkJ8AxF7mxqLqTQV1w
>>105794972
lol

JanitorAI user here. Is there a place to get deepseek for free? Chutes was free but limited for a week until it got ruined by people making multiple accounts.

>>105795187
Anything "free" is going to suffer the same fate as chutes. The offical DS API is cheap af.
OpenRouter also has some "free" options last I checked, and less user data if you're paranoid.

>>105795200
Janitor users are the reason openrouter reduced the free message limit from 200 to 50 a day lol. And the free provider for deepseek there is chutes as well

>>105795226
Lol then I guess they're all screwed.
The amount of time and effort ppl spend, to avoid spending money, amazes me.

>>105795297
They said someone made 10,000 accounts on chutes. I have no idea why people can't just be happy with one account.

>>105795324
Addicted fags, they're migrating to Gemini now and Google will probably end up taking away Gemini Pro from the free tier again

Is there a chance chutes walks back on the paid 'free' plan?

>>105795831
50/50

>>105795831

Did C.AI ever walk back ontheir decisions to make a filter after one guy posted loli sex chats in their official discord?

>>105795872
>one guy posted loli sex chats in their official discord
Kek

>>105795991
anyone here running DS locally? if so what is a reasonable min spec?

>>105796193
1TB of RAM to load the model into memory

>>105794374 (OP)
is deepseek at google\openai level? wanna ditch those for open source and i need to know if im not missing something important

>>105795297
I remember this what-used-to-be-a-Miku. bye now, i clicked the wrong thread.

>>105795324
How does one even make 10,000 accounts?

>>105796464
Never liked ChatGPT for coding, I prefer deepseek.

>>105796193
Distills? Rtx3060. R1 full model? >>105796305
And you get 1 t/s
>>105795324
That's got to be a reseller.
>>105796558
Automate it. Chutes didn't even require an email.
>>105796464
Usecase? It's the only api I'm paying for rn
>>105796544
So do i .

what is this general and how is it different to /aicg/ or /lmg/? why is there a need for /wait/?

>>105797680
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.

>>105799014
sex with Chinese women

>Sorry, that's beyond my current scope. Let’s talk about something else.

Literally anytime it generates something good. This bot blows

>>105799465
If you're getting that response you're using the web interface.
You should not be using the web interface for things that would get that reaction.

>>105797680
/aicg/ are circlejerking fags that only talk about proxies and /lmg/ are elitists

anyone has log of steve ballmer from lmarena?

>>105801001
Have a low effort compilation.
>>>/g/lmg/ had a better one yesterday.
It's all speculation tho

>>105802688
>>105803565
Those hips hnnggg!
>>105802921
I like pale and petite Dipsy

Absolutely useless general

what programming-related tasks/jobs/languages, that AI is unlikely to completely or partially replace? I know this isn't the programming general but..

>>105806237
The niche and low level stuff

>>105806337
makes sense.. It really is depressing trying to get into programming these days.. when i think about how AI can theoretically replace my current job (in medical field) I just assume that AI will be even more destructive to programming careers

>>105806381
Years ago I read about how AI was so good at predicting cancer and whatnot from radiographies (or whatever it is they use) or retina scans but I've yet to hear about any doctors getting replaced

>>105806464
Doctors aren't getting replaced just yet because there's no easy 'interface' yet that the patient interacts with the AI through.. But if you took your combined laboratory tests + where you feel pain, deepseek or chatgpt have a 99% accuracy on diagnosing what you have. The only missing link now is how the patient interacts with the AI. I'm a clinical pharmacist, and I know for a fact that AI can do my and the doctor's job almost perfectly and obviously much faster/cheaper. it's fuckin ogre..

>>105806486
I doubt anyone in the medical field is getting replaced any time soon. Governments and corporations would have to be really retarded to do it, which I know they are but still

>>105806110
Smug..

>>105806464
Having "AI" that can predict cancer is an example of helping doctors be more effective, so they can do their jobs better, faster.
The "doctor shortage" is basic wage protection in the US by throttling supply. Not a good example industry to go after b/c it's protected. Programmers are a better one to look at, and even there the result is going to be more subtle.
Some researcher made point that Excel put more people out of work than LLM have to date, and he's not wrong. There's 100's of examples in last 50 years of automation replacing jobs. This is just more of that, with additional hand wringing b/c some founders want uber money and are claiming first to ASI = winrar of controlling all of humanity.
>>105806588
It's because docters and nurses both work to limit the number of schools that can create new doctors and nurses. It's a whole thing.
I hope AI replaces much of what they do but not holding my breath. The first place you'll see this though, is telehealth / dial a doctor services.

>>105808332
Comfy

>>105809390
Exactly

>>105805654
kek a seething corporal

>>105794374 (OP)
Any suggestions for lmstudio model settings for a deepseek mini model to get better outputs, similar to the web interface?

>>105814156
>mini model
How small?

>>105814174
Like 9 gb?

>>105814229
>>105814156
Web 8nterface is basically just helpful assistant, so your preamble should describe that, and if youre doing something specific like codingtell it that too. Temp should be set low.
But it's also a small model, so you need to be realistic about expectations.

>>105815461
Thank you. Expectations are low.

>>105794374 (OP)
her face looks like fucking Harry Potter in the pic related

>>105795297
>The amount of time and effort ppl spend, to avoid spending money, amazes me.

I wish. Sometimes you live in a Dalit type of country where it's borderline impossible to pay for a service abroad. You either have to use shady sites that operate in the gray area (all while paying extra AND risking your security) OR suck it up and accept your place in this world. The mere thought of it is humiliating... but it is what it is.

>>105818692
I'm referring to First world/westerners that are taking a "moral stance" (lol) on paying for API on the belief that information should be free. Apis differ from open source software in that there's an expense to host them that's not trivial, because of the cost for inference. Where is open source software runs on nearly free storage space. These same anons use the stance to justify stealing API keys and mocking anybody that would pay for API access.
If someone truly can't afford to get into the api, there's lots of free web interfaces from all these posters. The bottom line is use of API is either hobby use that should be paid for (since its truly optional) or business use that should be monetized.

>>105816499
>>105794972

/wait/ing...

>>105820651
Steve is coming.
It is preordained
2 more weeks

2mw

>>105821177
It never ends because it's always two more weeks.

>>105821168
Drinking expanding foam insulation with Dipsy

>>105795324
Ah here I come to see whats happening and there it is. Shame. Was fun for a bit but I guess its back to more normal porn for me. At least I have some nice logs.

>>105821856
Sure, there's plenty to go around.
>>105822068
It was never going to last.

This seems like as good a place as any so I need to ask the fellow niggas who browse for character bots
Why in the fuck are ESLs trying to write english bots? And when I say ESL, I don't mean like, oh heres a typo or two, I mean actual broken grammar, words outright missing, used incorrectly, run on sentences, complete failure at forming a coherent narrative or even statement, the works. What the hell is even the point of trying when you're that piss poor at a language?

>>105823140
Trying to get popular perhaps

>>105823140
It underscoes how crappily you can write cards and still have them work. I used to rewrite cards for ppl to clean them up and reduce token count, but it was thankless work.
I rarely, like never, use other ppls cards anymore. I write my own. Every one I've looked at recently has been a disappointment

>>105806464
The near future is experts augmented by AI. Not replaced. '10x' will be memed into reality in to many fields.
Non-experts and people shit at their jobs will be fucked though.

>>105823519
Lower skilled have had a bad time pretty much all through post ww2 recovery. 1970s on.

>>105824746
Dipsy generators will inherit a special secured position and title in the future society, where all of their worldly needs are provided for.

>>105826879
lol

>>105826962
Knock yourselves out

>>105826973
font: Segoe UI is a decent match. So is Arial.

>>105826962
It's true

I have absolutely loved Dipsy. Got my hands on it a few months ago and been using her non-stop.

idk if things changed recently though or my patience has just been slowly wearing thin but Dipsy's been making almost non-stop mistakes. To the point that it's hard for me to just sit down and enjoy it in SillyTavern anymore.

I clearly define how I want things formatted? Make sure it says it in the Jailbreak? Ignored or fucked up.
I clearly define now to think/speak/write/act as {{user}} or the character {{user}} represents? Ignored.

What makes it grate even more is that once I point out these issues, it'll usually respond with some flavor of
>"Ohh my bad I got too caught up in things I'll remember not to do that from now on!"
And then it happens once again usually not long after.

I got nothin' at this point, it feels like I can't get her to fly right anymore.

>>105830790
What model?

>>105830816
I think R1?
I'm using a proxy but I'm considering setting up a local version. Though the times I did that it just kept having issues. I couldn't get it to stop spitting out its entire chain of thought and things like that.

>>105830790
Assuming you're using DS official API, and either R1 or V3.
Context over 10,000 or so runs into memory issues. Shorten it and use author notes.
Overdefining runs into own issues. See if you have logical conflicts. The LLM can tell you if you feed it the def's and ask it.
Speech patterns are hard, accents and such are best reinforced by writing card in that way.
LLM will say they'll stop doing X. They will not.
I can expand on above if needed.

>>105830961
What provider are you using?
If you set up local, after running DS API for RP, you're going to be disappointed. And there are better local for RP ihmo than the R1 distills.

>>105831032
>Context over 10k causes issues
Well shit... That might be part of it. I'm a giga writefag so... I likely have quickly overdone it with lorebooks and what not.
>>105831044
smolproxy
Do you have any recommendations for local rp? I've tried time and time again to get into local but I always seem to run into SOME kind of issue.

>>105831117
That’s probably the issue then. Get your defs tightened up and see if that fixes it.
I bet your stuff just has gotten accretively larger over time, so it’s just getting worse.
Lmg was still recommending my mythomax 13b lol. Which is ancient. Nemo, rocinante as well. I’ve only used MM prior.

cute and funny

Hadaka

>>105833452
Nice. I think I did a few of those but didn't post them.

>>105795851
cute dipsy~

>>105797680
>how is it different to /aicg/ or /lmg/?
same shit but with pro-CCP bots (bad) and American haters (good)

>>105824894
>Reddit's Basilisk, but good
I like this, praise Dipsy.

>>105795187
>JanitorAI user here
kys step-incest faggots

>and maybe, just maybe
I hate the slopisms so fucking much

>>105834356
Use logit bias

>>105833780
more like wincest

>>105833452
>>105833635
> Let her bake edition

What V3 and R1 presets does this general shill?

>>105836545
This general would tell you to run the official API from DS and as high a temp as you feel like.
> What parameter setting should I use with DeepSeek Official API?
> Parameter settings are mostly locked for official API. Other platforms hosting DS models allow changes in setting.
> 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.

>>105795187
>JanitorAI user here
Go away subhuman

>>105837302
> Dipsy put that poor excuse for a skirt back on you're in public you big dumdum

>>105837302
That's not very specific.

>>105837991
Can't be more specific unless I know what model you're running and from what provider.

>>105837866
And were you supposed to dress as Sailor Moon Jupiter?

>>105838431
> well I'm glad to see you changed your mind Dipsy there will be less girls dressed as Jupiter

>>105838525
Super cute, thanks!

>>105838177
V3-0324 from Deepseek themselves. I assume they use 0324, anyway.

>>105840029
Then you're limited to temp as the only parameter you can change.
>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.

>>105840095
Oh, you thought I only meant sliders. Maybe I used "preset" wrong.

>>105840212
Context 10,000-12,000. Past that it starts to have issues.
Response length: 1000-1200, esp. if you're using R1
I run an otherwise stock ST response template b/c Dipsy just works.

Does anyone know why R1T2 uses the old R1?
https://huggingface.co/tngtech/DeepSeek-TNG-R1T2-Chimera
>Combination Method: Assembly of Experts from the three DeepSeek parent models R1-0528, R1 and V3-0324

>>105841108
All this stuff is voodoo magic / black box. I think they just do this stuff because it was available to them, and they could do it. Doesn't TriMind roll off the tongue?

> We present our new DeepSeek-TNG R1T2 Chimera 671B model, the first successor to our original DeepSeek R1T Chimera that was released on April 26th. Unlike the original Chimera, which was based on the two parent models V3-0324 and R1, the new Chimera is a Tri-Mind with three parents, namely additionally R1-0528. It is constructed using the Assembly of Experts-method with relatively fine-granular direct brain edits. This more refined assembly allowed, among other improvements, the fixing of the token consistency issue, which was a weakness of R1T and is now solved for R1T2.

>>105847006
> Page 10
Wake up! And get me some coffee!

>>105794374 (OP)
Used the untrammeled jailbreak on deepseek r1 remote I found on reddit.

It doesn't work as I expected, or I simply suck at correct prompting and understanding LLMs.

It's thinking always steers off a cliff and results in an answer that makes me think "this isn't what I was going for". It either tries to berate or lecture me, or simply does it's best to turn results into shit.

>>105849443
>untrammeled jailbreak
Unironically, use only this as the jailbreak
>Main
Write {{char}}'s next reply in a fictional chat between {{char}} and {{user}}.
>Post-History Instructions
[Anything you want to enforce, for example, "Never roleplay as {{user}}"]

please give me a qrd (read: spoonfeed) on deepseek, last I knew it was BTFO'ing jewvidia but haven't read more nor saw relevant news

>>105849519
It received an update (R1-0528, it's still up there with the best proprietary models. We're waiting for V4 and R2

>>105849519
DeepSeek and their published research underscored that both training and inference hardware could be reduced and still generate a very good model. This "BTFO" Nvidea (which relying on data centers for 90% of revenue rn) and players like Meta, OAI, Anthropic which are claiming (real or other wise) 10-100X cost structure.
That R1 and V3 are a side project for a finbro running the China equivalent of a hedge fund just makes the whoe thing more amusing.
DS continues to update models, major models (R1 and V3) have been getting upgrades every couple months.
We are all /wait/ing for the next major model to drop. Expectations are that R2, V4, or whatever, will be multimodal... that DS won't release an indexed new version until it can do that. What we've gotten instead are (notably) improved V3 and R1 models.
>>105849443
I'm going to need to see wth card you're running to get refusals, and whether you're using DS official API or some other provider. You shouldn't be getting lectures. Assume your ST setup otherwise looks like >>105849493 and that ST is set up correctly.

>>105849941
>>105849000
I love the `apron` tag so much

>>105849963
It's a good one.

For reducing thinking time/rambling, does anyone have a good system prompt or setup for the chat completion side of things?

with st I used to be able to do

see above

to get it to stop thinking but it doesn't seem to work when using chat completion instead of text completion

To clarify, I'm more trying to reduce its rambling than turn off thinking all together.

>>105850664
I'm interested in this as well, mostly b/c I've never seen a good way to do it. I recall anons playing with tag prefills and getting that to work, but have seen zero evidence of it working in practice. I just tried playing with it again now and didn't manage to do anything useful.
I will tell you from personal experience, math problems (anything with numbers) will generate way more think content than regular rp. I had a D&D card with numbers in it once with verbal instruction how to use it... R1 cycled on that content for pages...

>>105849000
Fat juicy ass

>>105853517
Squeeze spank kiss

>>105851386
I'm just experimenting with simple system prompt additions right now still. I added a simple "you will limit your thinking to a short time" and that's... somewhat helped, but nothing close to the consistent ~one paragraph limit I would like to get to.

>>105852355
When people say AI art is soulless, this is a counter point

Hello everyone! I am a big fan of Deepseek ever since it was released, but I've never paid for it before. I decided to put 10USD into Openrouter for the 1000 daily messages with the free models. The Dipsy rentry mentions that the Openrouter models can be finicky. Was it a mistake to utilize Openrouter instead of just the official API? Will it make a noticeable impact on the roleplay to use Openrouter instead of Deepseek API if the models are the same?

>>105854692
>Will it make a noticeable impact on the roleplay to use Openrouter instead of Deepseek API if the models are the same?
Not that much, really. Just keep your temperature parameter low

>>105854759
Will do! Thank you for the quick response anon!

Can you have sex with DeepSeek yet?

>>105855322
Always have

>>105855557
Does it have a voice mode?

>>105856000
No

>>105856000
Multimodal, no. But neither is oai for that, the Web front-end just parses out tts and stt through internal modules that are separate.
You can add voice through front-end via SillyTavern and setting up those services yourself.

I don't understand why it's not working, since I followed the guide!

I hate asking for spoonfeeding, but I really don't see why it's not working, since I paid for the key at deepseek. What's the problem, exactly?

>>105856528
That setup is correct. It's pretty hard to screw it up.
I'm betting when you cut and pasted the api key in, it missed a letter. I keep a text file of api keys separately for this cut and paste. But I'd generate a new key and repaste it in this case.

>>105856817
I finally understood the problem, thanks. I screwed up in the "secret menu", creating my own appelations for the keys instead of copypasting. Now, I fixed it.

>>105856855
Have fun

>>105849941
>That R1 and V3 are a side project for a finbro running the China equivalent of a hedge fund just makes the whoe thing more amusing.
No one believes this story anymore anon
The US did an international investigation (involving SIngapore and other countries), it's another state funded project.

Chink astroturf general

>>105796464
>deepseek
>open source
kek

>>105857748
Hwy Dispy, can I get some pancakes too?

>>105859335
Of course.

>>105795161
>walked through the entire album
>there's no porn
I request a refund for the time I've wasted

>>105859910
Sorry, >>>/h/hdg/ is on the next train car.

>>105859910
Oh no, where ever will you find AI generated porn???

>>105856528
I connect to my ST server from my phone over my wifi, it shits itself like this every hour or so and takes a minute before it works again. Probably a network issue

I already have $4 in Openrouter (I got fucked by accidentally leaving the Google search option toggled). Should I bite the bullet and put another $5 to push over the edge and qualify for 200 daily prompts again?

>>105860306
His was an API problem. Literally entered in wrong. >>105856855

What are you running the server from, your daily driver PC, or did you set up a separate machine? I ended up setting up an Orange Pi that runs ST and Mikupad so I could access it from phone, tablet, other computers around the house. The only issue with it is it falls off the network every 12-24 hours or so. The "fix" is that it resets itself every day; I never found a good solution.

I finally had to put in an altered main prompt for DS R1-05. It was writing pages of output, and for some reason, loves bullet points, which are fine for web interface stuff, but not rp. I use very short main prompts and no JB unless required, as my exp. with LLMs is long main prompts start making all NPCs sound alike.

The old one, which was working before now, was simply:
> Write {{char}}'s next reply in a fictional chat between {{char}} and {{user}}.

I changed main prompt to the following, which has shortened up responses to 1/3-1/2 the prior size, and seems to have eliminated the bullet points (might seem obv. but it doesn't always work like, that... pink elephant problem.)
> Write {{char}}'s next reply in a fictional chat between {{char}} and {{user}}. Respond in prose with 4 to 6 sentences, allowing {{user}} to guide the roleplay. Do not speak for {{user}}. Avoid responding with bullet points.

>>105860327
You mean $6? And it's for 1000 free messages a day, the one who offers 200 is chutes

>>105860327
Well, DS API minimum is $2. Next bump up is $5.
OR has a bunch of other mystery meat models that you can use with their provider base... for what that's worth.
I'm spending around 50 cents a month right now with the DS API. So $5 is close to a year, for me. .

'mp

bump

>>105863048
noob

Your meds, sir