mirror of
https://github.com/deepseek-ai/DeepSeek-V3.git
synced 2026-11-05 00:47:48 +00:00
[GH-ISSUE #3] Everything is OK #1
Labels
No labels
closed-as-stale
enhancement
question
spam
stale
No milestone
No project
No assignees
1 participant
Notifications
Due date
No due date set.
Dependencies
No dependencies set.
Reference
deepseek-ai/DeepSeek-V3#1
Loading…
Add table
Add a link
Reference in a new issue
No description provided.
Delete branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Originally created by @chunyeah on GitHub (Dec 26, 2024).
Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-V3/issues/3
As Value99 said, since the model cannot support RolePlay, we can delete this issue.
Wish team success, thank you.
https://api-docs.deepseek.com/news/news1210
@Value99 commented on GitHub (Dec 27, 2024):
hey, just a heads-up, this model isn’t meant for that kind of stuff(you get what I’m saying, right?). let's keep it respectful for the folks who worked hard on it, okay? THX.
if you’re just chatting normally, you can try increasing the repetition penalty and temperature for better results.
@Andrey36652 commented on GitHub (Dec 27, 2024):
@chunyeah could you attach full request in json format (copy it from SillyTavern console)? And your generation params (temperature, top P, etc)? Although they should be visible in request too.
Some guys found out that incorrect format leads to severe repetitions. And that some parameters are better suited for creative writing.
Temperature 1.6-2.0. If you use really high temperature (1.8-2.0) it's better to adjust TopP to 0.98-0.95
@Value99 why your stance is so prudish? I think it's excellent that model can do diverse set of tasks: code, math, riddles or waifu RP equally good without being dumbed down by "muh, safety" and "muh, ethics"
@BboyMadara commented on GitHub (Dec 29, 2024):
Such a cringe stance, a generalist model should be able to do diverse task, including roleplay and creative writing greatly.
@jasonfish568 commented on GitHub (Jan 3, 2025):
Incorrect information. Deepseek does support role play. If you ever tried sending a prompt you will find it works.
@BboyMadara commented on GitHub (Jan 3, 2025):
Yes, but like the original issue state, it has very noticeable repetition issue, which for me personally the same as not working. Because after a while, it will start repeating line verbatim that already existed on previous replies. Even with a good CoT prompt, it doesn't really solve it, merely reducing it.
@jasonfish568 commented on GitHub (Jan 3, 2025):
I reckon you are right. It keeps replying the same thing even though I told it it’s wrong. I’ll do some adjustments on the params and see how it goes
@Happ1ness-dev commented on GitHub (Jan 8, 2025):
@Andrey36652
I don't use it for RP, but I also noticed it's pretty repetitive with writing tasks too. When I tried your settings, the model seems to just go crazy after a couple tokens. It only works if I set top_p to 0.3-0.7 when the temperature is 1.6+, but then it's repeating itself again. Increasing frequency penalty doesn't help on temps and top_p this low for me.
Do you have any advice?
@Andrey36652 commented on GitHub (Jan 8, 2025):
@Happ1ness-dev Do you use it with Deepseek API, or some other provider (Together, Fireworks, etc), or selfhosted?
Repetition loops also presented with tool usage. Deepseek acknowledged it and told they will try to improve it in the next iteration of the model, so let's hope.
Also regarding creative writing: I've tried to use Deepseek V3 Base model in Q4K_S quantization. It's completely different feeling comparing to instruct model. No typical "LLM prose" or "assistant slop" (I suppose deepseek overdid instruction tuning; they've mentioned it the paper that for RP and creative writing datataset generation they used Deepseek V2.5). But in exchange it's harder to steer the model where you want (as any base model though).
@Happ1ness-dev commented on GitHub (Jan 9, 2025):
@Andrey36652
Yeah, any other provider except official one. I just don't want my private data being trained on. One of the main reasons I use DeepSeek for coding and writing.
@Andrey36652 commented on GitHub (Jan 9, 2025):
@Happ1ness-dev other providers may have another notion of temperature. Also I've seen mentions about lower quality on other providers than official one.
@Happ1ness-dev commented on GitHub (Jan 9, 2025):
@Andrey36652 that's kinda weird, to be honest. Shouldn't they use the same model the official provider uses?
@Andrey36652 commented on GitHub (Jan 9, 2025):
@Happ1ness-dev they do, but Deepseek uses proprietary inference engine + cache
@Happ1ness-dev commented on GitHub (Jan 14, 2025):
I think I found a way to keep deepseek v3 from going insane on high temperatures. It's
min_p.Source: https://www.reddit.com/r/LocalLLaMA/comments/17vonjo/your_settings_are_probably_hurting_your_model_why/