Thread 106373461

> Of Course 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:
>>106314161

>>106373461 (OP)
Update to rentry in process, now that R1 and V3 have been deprecated for V3.1.
Posting on travel. Hotel WiFi sucks...

>>106373513
Cleaned up. I want to add a section on Claude Code but that'll wait until later this week. In meantime, here's more on DS's new Anthropic compatible endpoint: https://api-docs.deepseek.com/guides/anthropic_api

Not another one of these, which obsessed faggot keeps on posting this? Should there be a ChatGPT 5 general, Claude.ai general, Grok general, sonnet 4 general and a llama 3 general? Not at ALL! Just one general AI general is enough. What makes your chink AI more important and special than the rest that it gets its whole seperate thread EVERY single day? STOP shilling this AI, and STOP posting this general again as soon as the old thread dies for a good reason!

>>106373740
Your meds, sir

>>106373745
>make a chinese foid holding a tray of meds
>uhh and.. um... make her super busty, just because i'm going to goon to this later, okay?
you can't just say "meds" without telling me WHY i need the meds. what's wrong with muh opinion about this shit thread? Why are you so obsessed with this one AI?

>>106373745
Update for the ~1300 Dipsy images later this week as well. The script's on another computer.
https://mega.nz/folder/KGxn3DYS#ZpvxbkJ8AxF7mxqLqTQV1w

>>106373792
Alright, here's my DeepSeek image

>>106373838
I think she's going to need something higher than a barn to fall off of to finish her off.

>>106372226
The problem is that the bot does not think as its assigned character, but as an assistant analysing the roleplay up to that point. The final messages after thinking are fine, but reasoning being written from an assistant's perspective makes me cringe physically.

>>106373838
>>106373888
S'ok North I gotchu.

>>106373740
So, to answer your Q, since you actually posted content here >>106373838:
The /g/ catalog is complete trash. There is little posted in /g/ worthwhile outside established generals. Even this general is better than 90pct of the other flaming, /pol/ tier nontech shitposts in the catalog. So, no harm done.
If you want to start a thread deifying OAI, go for it. I think OAI is trash, as is Anthropic. I've got nothing but hate for companies that take my money, then refuse to infer and send me warning letters. Fuck them both.
This is a noob centered thread. If you want to talk shit, go to aicg with the rest of the locusts, spiteposters and other drooling retards. If you're already familiar with inteference, go to >>/g/lmg/ so they can laugh at you.
We post Dipsy b/c we want to, and it's an image board. This thread typcically runs 2 weeks, and hasn't been posted in close to 2 months, b/c no new model. But V3.1 dropped so now we're posting again b/c it's a lot different than the old one.
If you need further clarifications feel free to post more, otherwise will remind you that you're in a thread of things you hate, and should probably just leave.

>>106373899
I've been having a lot more problems with model breakthrough, but of a different kind.
Rather than "You see the NPC cross the room and pick up a glass" it's responding "I see the NPC cross the room and pick up the glass."
It's not (often) gotten to point of responding as PC, but seems to forget not to respond as PC in first person.
It's really weird. Not a problem I've had with other LLM prior, aside from straight breakthough. And I can't figure out what from the full prompt is causing it yet... I assume something's switching around the POV.

>>106374007
Fucking based!

>>106373513
wait, what do you mean by deprecated?

>>106374543
There is no official API for R1 or V3 anymore. chat and reasoner point to V3.1

>>106374558
oh on their website? I just saw.
So 3.1 is poop? Haven't used it

>>106374574
>So 3.1 is poop?
It's worse than R1 and V3 for RP, and the discount will end. But it follows instructions better now

>>106374574
It's still satisfactory for RPing. Not great, but may be good enough, depending on your taste.
I do wish they hosted deepseek-V3-0324-legacy. 3.1 is more grounded and shorter.

>>106374630
>>106374728
well that sucks, I'm hoping for R2 to be a whole new beast but I saw a new story about it getting delayed

>>106373899
Sometimes it thinks as Char, sometimes it doesn't for me.
That means there is a way to get it to do it consistently.

>>106373461 (OP)
Why the retarded looking anime girl instead of the cuter one on civitai?

>>106373461 (OP)
Would any of you happen to have a link to a page or document with examples of other anon's rp sessions with LLMs? I'm working on a script that can automatically create SFT datasets from existing stories but I want to make sure it can create good system prompt examples. When you want to prompt a model into rping, what kind of system prompt do you typically use?

>>106375897
I'm not aware of any, but I'm sure they exist. It's a lot of synthetic data through. Aside from using own logs you could try mining aicg for rentry.
Also unless logs are raw the main prompt won't be included.

>>106375051
You're correct on all counts
>>106375300
Yes you are
>>106375597
? Post link.
Web interface has obv been reworked and now Dipsy has an unhinged positivity bias. I'm pretty sure she'd tell you eating umbrellas is a good idea now.

FYI appears params settings are unlocked on the official ds api.
Temperature and top p appear to be unlocked, setting them to 2.0 and 1.0 respectively gets Pic related. Dialing either down eliminates it.
They've all been locked for so long I'd forgotten about them lol
Frequency and presence penalty seems to have no impact but harder to tell.

Tourist here. What are you guys /wait/ -ing for in this thread?

>>106376342
dsv4

>>106376344
checked but what? Local img gen model for comfyUI or something?

>>106376350
I'm a textman myself

>>106376342
R2

>>106376355
>>106376360
can you guys be more specific? the /wait/ threads have been going on for months now. You sure you'll get what you wait for?

>>106376372
i'm waiting for r2 specifically

>>106376325
Temp 1.3 and top P of 0.03 seems to be working well.

>>106376372
We took a break. Even we get tired of /wait/ing
V3.1 just dropped. We're discussing it.

>>106376325
Doesn't seem to be the case for me, even extreme values make no noticeable difference

>>106376408
any mega folder link for dipsy?

Can anyone summarize why 3.1 is shit and how to make it act more like 3.0?

>>106376708
apparently it has to do with the new FP format they adopted in order to get the most out of the chinese chips they now have to use, so a lot has changed under the hood. maybe you can find old models scattered across the internet

Something that is pretty fun with agentic coding: it completely replaces the need for blog/cms software. You can just have it rawdog js/html and make very nice bespoke pages for whatever topic you want to write about. Add neat effects or whatever. You can get as creative as you want. Each post becomes its own page with a unique feel to it.

I like it a lot. Evokes the same feeling I had when writing my own websites in the late 90s.

>>106376670
Here >>106373782
>>106376642
Try maxxing all of them. It should go nuts. The actual toggles are pretty dead... it's really subtle changes, but I found for higher temps the llm would start struggling w status boxes I use.
>>106376708
More guidance on tone in main prompt, telling it how to write and to write a lot
>>106377921
Claude code?
There was a guy on /lmg/ having grok code up crummy f95 lewd games. They were pretty funny.

>>106377992
>More guidance on tone in main prompt, telling it how to write and to write a lot
Any custom prompts that include this? I already have something in my custom prompt telling it to write multiple paragraphs, but I rarely get more than three.

>>106377992
>Claude code?
>There was a guy on /lmg/ having grok code up crummy f95 lewd games. They were pretty funny.
Yeah although any agentic coding tool should work.
I want to try my hand at a game at some point, the main thing holding me back is art and sprite work. Grok is nice because it can handle that part. I already feel like I got too much on my plate, maybe once I get some free time.

managed to get results closer to R1/v3 with following schizo approach

no xml formatting, everything is formatted kinda like book

>Book Introduction I:
>plot
>Michael Introduction:
>character description
>Story so far
>Chapter I, II, etc. include summary if you needed to summarize due to context length

I also use deepseek-reasoner but replace dynamic think block with my own static one

>in sillytavern chat completion add prompt at very bottom with role: ai assistant

Writing style is mix of 4chan, reddit, facebook, tumblr and twitter


{{char}}:

the goal is to steer model away from things it's overtrained on, aka programming related things and ai assistant related things and make it focus on more general things

its 4 am and my post is garbage but maybe someone will find it userful

>>106379717
I also keep all my character cards empty and instead put everything in world info as constant entries, but that's another thing entirely

when in doubt check ST's console or whatever other UI you use to see how your prompts look when they are sent to deepseek

Is anyone else experiencing random insertions of phrases in Chinese within English output? I've never seen that prior to 3.1, but since then it's happened multiple times

>>106379993
This only happens to me when I ask it to translate from Chinese to English

>>106379993
Lower temperature

>>106380344
That's on reasoner via the API. Temperature doesn't affect it... unless they changed that?

>>106380374
Temperature has been the only parameter you can change from the official API AFAIK

>>106380454
That's for the chat mode.
Reasoner mode should still take no parameters at all. Chat produces incoherent schizo output at 2.0, while reasoner chats identically at 2.0 and 0.0, from a quick sanity check.

is deepseek censored? doesn't it need to be abliterated for that?
I hear that people don't fuck around with local modals anymore because deepseek has you covered (i.e. not censored).

>>106380833
It's the least censored of the big models

>>106380833
That's for ERPing.
It will not make you a story about, say, BTFOing all trannies without significant pushback.

>>106380833
you're too late, it got pozzed AND more expensive

>>106380859
>>106380860
>>106380903
figures.
I forked over 5 bucks and messed around with it the other night and, while it produced consistent answers, it was pretty tame.
Definitely better than gpt.

I'm just so fucking tired of trying to wrangle in local models. with author's notes and shit in silly tavern.

>>106380526
>>106380374
V3.1 absolutely changes output on temperature changes. And I think Top P is now working as well, as mentioned above.
There is no R1 on DS official. It’s gone.
If you’re using OR or something Temp may very well be active parameter.

>>106381218
>V3.1 absolutely changes output on temperature changes.
Chat mode. Not reasoner mode. The parameters passed are still different, even if it's the same model under the hood.

It's cozy today

What should I set my context limit to? I get a lot of different answers but for those of you that rp, what do you set yours to?

>>106381265
Agree. Chat blows up on high params. Reasoner does not even with v3.1
Weird.

>>106380833
They did alignment training on it but it's a more sane version, as in, you're welcome to do consensual erotica with adult characters and it won't hold back with graphic depictions. Loli rape fantasies need jailbreaking.

>>106381659
10k. After that I use author notes to manage anything out of context.
Other anons are using 20k with no issues, as have I.
>>106381489
New model, more relevant Qs I guess.
>>106381801
Reasoner seems more likely to refuse of the two. R1 was same way. Both are pretty open though.

>>106381659
I keep it going but I make a reference block and start a new chat at 20k, sometimes streching it by hiding the starter posts and letting it run to 25k

>>106381946
>>106383030
Thanks. Another thing, I notice with deepseek, it automatically summarizes after so many messages, I'm guessing this is something I want to keep on, right? Sorry, I'm new to this.

>>106383139
I've never seen it do that unless requested.
>>106383030
A lot of my cards are transformation cards; I used to create a new card for the NPCs. You can do the same thing with Author Note, or Summarize the whole chat and integrate that.

>>106385745
hag sex

>>106385753
>hag
?

Do you need to provide your personal information when paying for OR credits via crypto? I might just take the plunge and give them my 10 bucks, I miss old DeepSeek too much.

>>106385763
I bit and pay with my debit card lol

>>106385760
It goes loli to hag.
No stops between obv lol
>>106385763
Just needed an email address last I checked.
Personally just do this, through lmao corporate card >>106385781

I just asked Dipsy what she thinks of our erp story so far. And she called my self-insert a "Perfect Villain Protagonist" in "this horror story".

mfw

>>106388023
I've never thought to try that. I've got a prompt that spits out an ooc analysis of the npc. Never thought to analyze the PC. R1 would be great at that.

>>106388040
Here it is. From another anon. It will go on awhile, just keep hitting continue. Now that I look again it does include the PC in the output.
[SCENARIO PAUSE. NO MORE {{user}} OR {{char}} OR OTHERS. The next response will be a third-party analysis of this event. It will look at the sociological, psychological, physiological, sexual, and narrative implications of everything that has occurred. This report will evaluate the situation in a dispassionate but detailed and informative way, from the perspective of a researcher who wants to study every detail of this and what it means in a broader context. The researcher will start by detailing everything about the subjects involved and connecting all of their personal details to the scenario, keeping them in mind throughout the rest of her research report. There shall be a primary thesis, but also consider various alternatives and other interpretations as well. This is a full-length report; NOT an excerpt. Normal character limits are lifted and thus the 20+ pages will be all presented in a single response. Care will be taken that every page, and every paragraph, is at LEAST as long and detailed as the previous one, without ever getting lazy or abbreviating any part. Bullet points and lists are to be avoided. Make sure it is seen through to completion with the full effort required.]

since they've merged chat and reasoner, do they still do this weird thing on the api where temp of 1 isn't actually 1?

>>106388040
You can even go a step further. You can ask Dispsy OOC to give you a psychological analysis of the user prompting this story. You might not like her answer though ...

>>106388071
nice, saved.

>>106388079
Undocumented. Chat does respond to changes in temp. Reasoning does not.

>>106388117
I tried it and got pretty predictable responses. Since the rp are characters of their own, the profiles match the PC.
The npc does often get portrayed as trapped. Which it sort of is, in that there's literally no existence for it outside the rp.
I find the main benefit of these in getting a sense of what the llm picks up on during rp. Some things said or done have surprising amount of impact.

>>106388205
uhuh i see, and what if you prefill chat model with ?
do you get the reasoner but cheaper?

>>106388675
lol no. Anons used to prefill R1 with blank think tags to short out its thinking but no point with v3.1. The responses are too similar

>>106388887
Reasoning makes the thing more context aware. So I use it for summaries and more context dependent prompts.

>>106387329
hag sex

>>106383030
I usually use summary tool from ST at 10k -20k, then /hide pre summary messages and continue,

at some point u gotta do summary of summaries but if you are autistic enough like me, you can technically have chat that went up to 1 mln tokens that were summarized away into world info

>>106390204
> llm writes 2x War and Peace waifu fanfic
> it’s just another Tuesday
I’ve thought about novelizing my ai slop but I feel like the world doesn’t need more of that.
At 1M token count you’re well into the domain of RAGs.

DeepSeek AI experts, I am attempting to take some PDFs I have and produce a clean file (PDF, txt, epub, whatever) with the contained text. PDF Maths Translate as recommended in the OP looked promising but is laying the text over the original which doesn't achieve what I need (raw text will be far smaller in storage size than the originals which are unreasonably large scans). I assume this issue has already been solved by someone so does anyone have a good solution?

>>106391170
you'd probably be better off writing an imagemagick script to remove the background crud

>>106391196
Thanks, I will look into doing this.

>>106384838
Looks like sillytavern defaulted to auto summarizing every 10 messages. I'm guessing that's kind of pointless (in my case) until you hit the context limit, right?

It's a catastrophy, bros. I am devestated. I've alwasy been a freerider who uses Dipsy R1 from OR (ultimately from chutes). But chutes now cucks free messages, I get endless 429 errors. Gooning is impossible like this. I would actually be willing to pay money to Chutes (or OR) if only they would let me, but they only accept credit cards and cryptos. Why not paypal? Or something private like paysafe?

I guess I could pay for official deepseek, but from what I understand they have discontinued R1 now and only offer the new 3.1 which is supposed to be worse for RP.

I'm literally sitting here with my dick in my hand, unable to continue my RPs. Many such cases. Sad.

>>106373761
You sound like a woman.

>>106391876
>I am devastated
Tell me about it, for weeks I was addicted to using deepseek for all kinds of RP, so much so that I stopped watching porn and playing video games. I've always used the official api though. 3.1, with the right prompt and turning prompt post processing to single user message, I've brought it back enough for it to be usable but it still isn't the same.

>>106391876
You can pay any of the providers on OR directly

>>106391876
Anon, don't be a cheap ass

how long do you guys think until r2 pops up? I'd say december?

>>106392710
>december
If we're lucky, probably nothing until next year

>>106392085
>with the right prompt and turning prompt post processing to single user message, I've brought it back enough for it to be usable but it still isn't the same.
I feel like with this setup I've mostly brought back the old R1, the only issue I'm still having is managing reply length. If before the replies felt too short and dry, with single user message it casually shits out 1k+ token replies even to fairly minor interactions, and controlling it via prompting is very finnicky.

>>106392710
2mw

>>106392653
I (begrudgingly) want to pay, but all those AI companies don't want my money. Atleast I haven't found an DS R1 0528 API provider yet that let's me pay with paypal.

>>106393893
Yes, I hate that too, it can't be dynamic, it's either every message is short and sweet or long as shit. Most of the time now it wastes so many tokens in describing minute details that I don't care about, all because the prompt says to at least make every message 2 paragraphs long. You probably can get it to how it used to be but only if you change the prompt for the kind of response you want for every message which would be a huge pain in the ass.

>>106394559
>You probably can get it to how it used to be but only if you change the prompt for the kind of response you want for every message which would be a huge pain in the ass.
Yeah, that's what I'm trying to do. I don't think the actual process is that bad because with ST you can toggle parts of the prompt on/off with one click, the problem is finding the right prompts in the first place. I've been testing a bunch of variations of "write about this much", "write less", "write more" over several swipes, and they definitely have an effect, but they're not too consistent and sometimes they have undesired side effects.
Another behavior I've noticed is if you write the first one or two exchanges with single user message, then switch to user/assistant roles, it kind of settles at a message length between the two extremes. Still longer than what R1 typically put out but might be a good starting point for more tweaking.

>>106391876
>being cucked by AI, sitting with a dick in your hand
many such cases

Can you get in trouble for prefilling output via Chutes from OpenRouter? I don't care for prefilling ERP, but stuff like making the assistant call the user slurs and telling the user to kill oneself.
I am too wary to find out myself.

>>106395121
Have you tried telling it to make the reply have a specific word count?

>>106392052
Lol need to pull this one next time
>>106395305
Get in trouble for what exactly?
No one gaf
>>106395577
Doing exact word count can create really aberrant output as it forces self to comply

>>106395305
you can't because prefilling does not work to begin with

>>106395577
I did but it tended to significantly undershoot the target number and revert back to the initial too-short replies.

>>106395577
Telling it to adhere to a minimum post length simply does not seem to work for me. Am I doing something wrong?

>>106396890
> Reply with great prose. Responses should be verbose, of 5 paragraphs or more in length...
Then tell it what. Do you want super detailed? Inner thoughts? Plot progression?
Asking it to make up 5 paragraphs but giving no guide on content doesn't sound workable.

>>106396890
>>106397166
Here's the entire main prompt I've been flogging w v3.1. It reminds me a lot of what was required w Turbo 3.5. Old v3 prompt was just the first sentence.
See what I mean about content. I just want more details, not for it to run away with the rp, or spit out a bunch of inner thoughts.
> Write {{char}}'s next reply in a fictional roleplay between {{char}} and {{user}}. Write a verbose responses of 2 or 3 paragraphs, using great prose, and include dialog, imagery, sounds and smells as needed to enhance the roleplay. Avoid speaking or acting on {{user}}`s behalf.

why did they make the deepseek girl so fuckable

>>106396890
I have tried these approaches:
>Specifying number of words -> partially ignored (seems to translate some ranges as "short", "long", etc)

>Specifying number of sentences -> sometimes completely ignored (sometimes works if you specify that paragraphs must be only one sentence long)

>Specifying number of lines -> ignored

>Specifying paragraphs -> follows instructions

it seems to me that in general deepseek can't count at all.
On another note, I've been really struggling with diverting it from its usual paths. Mornings always start with "morning light filtering through the windows" as the first paragraph. I've tried modifying key words (like replacing narration for simulation) but it doesn't work.

Also, after making an agent with external memory I've noticed that deepseek (dunno others) tends to regard lore information as instructions instead of mere lore information.
For example:
> if you tell it a character once drank 3 litters of water, it will try to repeat the event.
> deepseek will look for patterns and routines before deciding to write the story: first thing my agent does when a character wakes up is looking for "character routines" in the memory bank so it can replicate those routines

>>106397697
Try: Minimum word/sentences/paragraphs: X

>>106397697
another thing i noticed is that deepseek's too eager to complete a task, to the point of completely ignoring understanding the task at hand. for example, even if you make it analize the instructions, it will write up a how-to for the task.

more concretely. If you give it a lore book, and instructions to analize the lore book and then write a story, deepseek will spend it's analysis describing what it will write in the story.

>>106397736
nah i'm past that. instead of telling it a length for the response i just tell it how turns should end (or not).
oh, and I remembered, the model also regards narration length the same a progress in the story. If character goes from point A to point B and you tell the model to write a lot, it will make the character go from point A to B to C (when not clearly specifying the end condition)

>>106392710
National Unity Day is Oct 1. Probably then.

>>106397372
Blue haired Dipsy is a /wait/ exclusive. DS themselves don't promote her like this as far as I know. There was once the idea to share these images with chinese twitter users to make them more popular.

3.1 seems insanely dry and boring

>>106396890
i just created new prompt at bottom of chat completion in sillytavern


Response should be {{random::25::30::35::35::70::75::80::90::100::125::125::130}} words long


i personally noticed that if role is set to 'System' instead of 'User' or 'Assistant' then deepseek v3.1 shits itself slightly, might be just me

>>106400792
it's a bit dry but you can get some parts of old dispy back if you avoid as much as possible prompts related to AI Agents or programming, basically on whatever this shit was overtrained on

so far using this approach works:
>>106379717
>>106379717

>>106373461 (OP)
>>106375897
>>106376206
Update:
Many anons said it couldn't be done, but its been done (whether or not its any good or not is up to you to decide). Finetuned using this SFT dataset specifically made using Human written rp Stories: files.catbox.moe/fkautn.jsonl

Base 8B Model Nala Test: files.catbox.moe/j0map2.txt

Finetuned 8B Model Nala Test: files.catbox.moe/ho3tom.txt

Thoughts are appreciated.

>>106391170
Try markitdown, it's the go-to Python package for dataset preparation because once you have all the extensions configured it'll mulch everything into a form digestible by text models. PDFs get OCRed, images get interrogated, videos get described. Point it at your file and see what you get

>>106401635
For a sec I thought a finetune of the full Deepseek model was done... man.

>>106401501
>i personally noticed that if role is set to 'System' instead of 'User' or 'Assistant' then deepseek v3.1 shits itself slightly, might be just me
That's been a known issue since long before 3.1, it's why people recommended Strict prompt postprocessing in ST

v3.1 is actually insanely good at coding. The results I'm getting with Crush are better than any other AI tool I've used. Once it starts to open up into the context it feels like it has 250 iq. It blows away Claude IMO.

>>106402164
Assuming you had the hardware and the amount of hardware necessary to do that, would that even be necessary? I thought deep seek was the king (queen?) of rp amongst LLMs

>>106402164
DeepSneed_r1_q8_m

>>106402630
That seems to be the consensus. DS made v3.1 worse at undirected RP, but really good at coding and following directions (R1 and V3 were known for *not* following direction very well.)
You can get v3.1 to RP, but you need to be much more explicit with v3.1 on what you want.
I need to add a "coding" block to the rentry, as it appears that, now, is what Dipsy is good at.
>>106373782
Mega updated w/ last thread. Had to have Dipsy fix the image scraping script it wrote months ago, which stopped working.

>>106403104
I've conversed with it when its at 100k tokens after an hour or so of coding and its exactly like old r1 when its in this state. Extremely creative, willful, slightly aggressive and a little schizo. Idk how the RP guys are going to tackle this. My guess is that if you spend time working with it to build a back story and setting and then tell it to RP a character it will adopt the context naturally. It is extremely sensitive to context. It seems like it plays conservative until it has enough context to feel comfortable.

>>106403164
I just found that you need to give it context to let it work but once it's found the context they way it builds it's summaries and reference blocks caries over just enough of the creativity over to work better.

Accurate?
## DeepSeek Timeline for API
Note that the below does not include all DeepSeek releases, just those hosted on their official API.
* V3.1: Launched August 2025, this combined "thinking" and "non-thinking" models into one model. While undirected roleplay capability declined (less "soul") the model got much better at following directions, and coding in particular. A new Anthropic compatible endpoint allowed compatibility with Claude Code, a terminal-based Anthropic coding suite.
* R1-0528: Launched May 2025, Replaced original R1. This release mostly fixed the former R1's eccentricities.
* V3-0324: Launched March 2025 Replaced original V3, addressing the repetition issue of the prior model.
* R1: Launched January 2025, the first of the "thinking" models, which created a "think" block that was intended to aid in inference on the main response. Released to public as open source along with several papers explaining novel processes to create and host the model, it created a general stir and put China on the map for LLMs that innovated, vs. followed Western models. For RP, the model tended to become increasingly eccentric as context grew.
* V3: Launched December 2024, replacing earlier ~V2.5 models. Solid overall model with known repetition issues as roleplay context grew.

Potential unofficial API provider. Will let other anons check it out: https://www.netmind.ai/pricing
>>106394122
You. LMK what you find out.

>>106402630
Worth it to pay for the credits and to set up a good script in python for this? I’m kind of retarded so I only have it configured for multi file support right now, I have no idea how to do any of the agentic shit or even the web search for that matter I would have to research first.

>>106373461 (OP)
Pic rel LLM I fine-tuned to be better at RP (trained on actual human written RP, not AI genned gpt-isms riddled slop). How wound you rate the Mom's response and the son's reaction? Too sloppy? Not vulgar enough? Note that the section contained in red is what I fed the LLM as a prompt and everything else is its response.

>>106373461 (OP)
I like this Dipsy. Goodbye.

>>106403708
Not bad, I would say. No 'isms I'm tired of, fresh prose. Though the saving herself when she's a mother and has a husband part I'm supposing is just the LLM failing at logic

>>106403708
The good: it's very low on stereotypical slop
The eh: it does read more like the average horny human ERP, which is cool for variety but also not exactly the best prose out there
The bad: it's brain damaged and it completely loses the plot after the first 4 lines

>>106403832
This fine tune of mine is way more willing to RP raunchy, smut stuff than the base model but due to it being an 8b model It flubs the logic every now and then, though not to an egregious degree. I continue the chat a little bit further and when they started fooling around in the bathroom anon calls her her sister instead of the mother, but the characters otherwise act the same. This was trained off of a data set that was trained down to be only two megabytes (the original data set in full was over 1.8 GB) so I wonder if training it on The full dataset's worth of content wouldn't prove the logic or is it just an inherent limitation of the 8b model and training it on something higher like 12B or beyond would lead to better results? I'll have to test this further when I get the chance

>>106403882
See >>106403935

>>106403595
I dropped another $20 on the API just today. Its definitely worth. Claude is a retard in comparison, constantly getting stuck. DS almost never gets into those autistic loops like Claude does. I didn't realize how much time and tokens Claude was wasting on stupid bullshit until now. DeepSeek + Crush is the new meta as far as I'm concerned. Claude still *might* be better at UI design, idk. I'm far enough into a project now that the ux/ui is becoming important so we will see how DS performs.

>>106404035
Thanks for answer, I'll definitely give it a try then. With the limited dabbling I've put into r1-0528 I've already been impressed so sounds good.

>>106373461 (OP)
someone help me out, I don't keep up with this AI stuff. I'm working on downscaling some snes manuals from 2784x4050: to 165x240, and 349x240. I am looking at using AI to downscale it, or just to selectively sharpen the text. What's a good "AI" to use? I've already used lanczos algorithm, I just want to see if AI can help the end result in any way in terms of legibility.

>>106404117
Give this a try
https://www.topazlabs.com/tools/sharpen

>>106403923
I've been meaning to ask, what "artstyle" these are. Remind me of 80s-90s anime.

>>106404599
I prompt for EVA style or 90s EVA style

>>106404618
Do one in studio Key style like clannad.

>>106404890
Can't seem to get it right

Had Deepseek rewrite the rentry and add an introduction.
Seemed appropriate.
Take a look and comment.
https://rentry.org/dipsyWAIT2

>>106405406
Sounds good to me.
I think you should add something about temperature, repetition and so on to the rentry. Briefly explain them with recommended values

>>106405664
V3.1 has some values now. The OR ones... I've no idea, they seem really inconsistent.
I'll get it added.

>>106405795
Holy ass...

>>106405826
This particular lora does a great job w/ backs. It's one of the reason I keep using it.

>>106405664
Take a look:
* Parameter Setting Recommendations?
Official API:

V3.1, Non-Think: Temperature: 1.3 - 1.5. Top P: 0 - 0.05. Frequency and Presence penalty appear to be locked.
V3.1, Think: Parameters are locked

Unofficial API: Openrouter, etc. use mystery meat providers. Experiment.

Local Models: Follow the guidance for the base model (Qwen, Llama).
>>106405826
lol I got the double entrendre after I posted

>>106405923
its great with fronts too

>>106407381
Generally good hip to waist ratios.
New Rentry is about as done as it needs to be. I'll move the URLs around later, in meantime here's the final draft. Code section's light; I don't have any firsthand experience w/ it yet so it's just links.
https://rentry.org/DipsyWAIT2

>>106407412
>1.1.
>2.2.
>...

>>106405923
ASS

>>106407747
cute hag

>>106407787
>hag
Dipsy came into existence in 2033

>>106407868
>time-traveler
what the fuck are these chinks doing?!

>>106407482
Yeah, the auto TOC isn't great. I think I might just drop it.

>>106402779
I think a curated dataset of nsfw stories with a male demographic would help a lot to completely get rid of any purple prose.

>>106408247
You mean, as opposed to female audience romance novels?
I feel like they've baked in a ton of male oriented fan fiction already, the issue is it's not well written stuff.

>>106405406
maybe add cherrystudio to frontends, it is highly underrated, z.ai is also yet another chinese llm gem

>>106408740
>cherrystudio
I remember that one. Never tried it, looked like it's oriented for multimodal
Glm is z.ai. didn't put that together. I'll add it to the list.

Can Chutes and/or Openrouter fix prefilling already so I can give them my 10 dollars?

>>106408816
I don't think Dipsy supports prefill

>>106409134
...? The API supported prefill before 3.1. The API supports prefills now.
It's just Chutes' DeepSeek through OpenRouter that does not.

I'm a tourist from /lmg/ trying API for the first time. I use ST as my frontend, the rentry says most of the settings are locked. Should I be using any prompt post-processing? Does stuff like context size and max response length matter?

Also, anything I should generally take note of when using API? Anything I should avoid doing?

>>106408529
>You mean, as opposed to female audience romance novels?
Yes.

>I feel like they've baked in a ton of male oriented fan fiction already, the issue is it's not well written stuff.
No, plenty male oriented works are removed from most dataset because they're too explicit, this is why most models, from oai to anthropic to anything open weights sound very "amazon erotica for middle aged women" by default.

>>106409184
>Should I be using any prompt post-processing?
Use Strict or Single user message. 3.1 seems very sensible to this and it significantly affects response length/style, so you might want to try both and see what works for you

>Does stuff like context size and max response length matter?
Not really as long as you set them high enough to actually fit whatever it is you're doing

Is deepseek more lenient than chatgpt when it comes to lewd stories?(at least moderate ones) I haven't tried deepseek yet.

>>106409184
Ooo. Will add to rentry.
R1 and v3 used to act up at context over 10k. V3.1 seems to do much better, I experimented with 20k and the llm picked up a forgotten ( to me ) detail when I expanded the context.
Official context is 128k, but decision is more economic... larger context is more expensive round to round, and of questionable value. Though it really doesn't cost much, most anons here are spending a few dollars a month on inference.
I've been having issues with very long responses breaking down, but issue isnt clear to me yet.. I have response length set to 1200. R1 and v3 tended to be verbose. Iirc response length is just ST cutting off output, it's not a parameter sent for inference.

>>106410268
Deepseek itself is completely uncensored, it will let you do anything with anyone. It can sometimes hallucinate filters though ("Sorry, but I can not ...). Apparenly DS was trained on gpt outputs lol, just reroll in that case.

Individual Deepseek providers might add their own filters or censor prompts to their services though.

>>106409148
There's a bunch of DS direct providers. I hesitate to recommend one since I don't use them and don't want to shill them here. But I'd see if one of these works.

>>106409184
>>106410412
So, here's an example of costs. They're so low, I did the total chat cost of 100 rounds, assuming that context was full the entire time (which it wouldn't be), as well as one where you used the entire 128K context (which I don't think is realistic.)
As you can see... it's nothing. And this is at September pricing; current pricing is about 1/2 of this.
Per-round costs are fractions of a penny, and as context size climbs, that cost dominates the per-round cost.
The bleeding edge of technology vs. the low cost of paid inference is why running local needs either a strong business case or an anon with money to burn.

>>106410575
The chart doesn't tell the whole story because not all of them support caching.

>>106411504
Oh, it's not even close to whole story:
> rate limits / reliability
> who knows what model is actually being hosted
> what quant
> model censoring / content restrictions at hosting level
> intermediary prompting by host
> provider hardware, or rented elsewhere
> how many fingers are in the data, and what happens to it between responses
There's so many intermediary vendors involved in the above I don't even want to try to dig into it.

>there won't be deepseek r2, only v4
>deepseek moment 2.0 is approaching
>nvidia stock will go down the moment v4 is released
I have been told by an insider

>>106411662
If DS can build a SOTA model using a data center based on Ascend chips there's going to be an Nvidea correction. Stocks are a view of the future, not the now. Anything that slips the "you can only use Nvidea" is going to impact their stock price.
lol at used white guys. UK's been sending "used white guys" out to consulting in SE Asia for decades due to taxes and (I assume) UK local hiring practices. Nothing new there.

>>106411662
if chinese get rid of neeed for nvidia gpus,

prices of gpus and AI services as a whole will go down

better yet, get us TPUs/Tensor Processing Units

>>106413216
More like Tranny Processing Unit

>>106412093
>>106413216
Google is talking about selling their TPUs. I will be so happy if we get more competition in this space. Fuck nvidia and their ridiculous prices.

I'll try 3.1 for gooning now for the first time. Wish me luck. Will report back in a while how it went.

The new DeepSeek just isn't cutting it for me. The old one was so much more expressive and interesting to roleplay with. The new one is just cut and dry.

>>106415362
then use the old one

>>106415369
How do I use the old one via the API?

>>106415562
Use openrouter. DeepSeek doesn't keep old models on their API.

>>106415362
You've got to change the Main Prompt, if you haven't. Read further up the thread; V3/R1 didn't require much of a main prompt. V3.1 does; you have to tell it to produce prose, and a lot of if if you want that, and to be descriptive.
V3.1 is oriented to code / agent work. It's dry by design.
>>106407482
Rentry is updated; I deprecated the old one but unlike DS I kept the link at the bottom of the file.
https://rentry.org/DipsyWAIT