discussion Good LLMs for Eroge Translation

[planB]

I'm in the process of translating some japanese eroge into English. But my translation and formulation skills are limited so I use common LLMs to translate and adjust the narrations and dialogues. But due to the content restrictions I, more often then not, run into the problem that the models won't translate or formulate the texts for me. Thus I explorered unrestricted offline LLMs but I haven't found any that reaches the level of quality in this area quite like GPT-4.
Does anybody have any recommendations on models I could test out?

 

[faplordship]

Grok it is uncensored

 

[enderman118849]

prompt injection can trick gpt4 into processing explicit material

 

[[丂ㄖ尺卂]]

I'm in the process of translating some japanese eroge into English. But my translation and formulation skills are limited so I use common LLMs to translate and adjust the narrations and dialogues. But due to the content restrictions I, more often then not, run into the problem that the models won't translate or formulate the texts for me. Thus I explorered unrestricted offline LLMs but I haven't found any that reaches the level of quality in this area quite like GPT-4.
Does anybody have any recommendations on models I could test out?

just running it in translator++ and manually fixing the gender terms works really well for (mtl)ing visual novels and japanese RPGmaker games

 

[planB]

just running it in translator++ and manually fixing the gender terms works really well for (mtl)ing visual novels and japanese RPGmaker games

The problem with this is, a lot of character identity and quirks are lost and with a good LLM, this can be preserved to some extend. And since I'm in contact with the actual devs and it's possible that this becomes an official translation, I want a higher level of quality.
(And yeah, I know the hypocrisy of trying to be an official translator without having fully learned the language and using digital translators...)

 

[gigadigga]

For context: I'm a translator JP > EN.

There are none for proper translations. If you're fine with a general, often inaccurate gist, then pretty much any LLM/machine translator software would work.

However, if you can't read Japanese, you shouldn't be machine translating. It'd be impossible for you to understand what's lost in the process, wrong, or omitted. Not to mention, as you pointed out yourself, you usually lose all nuance, character, tone, personality, humour, etc. As far as I know, there are no machine translators that are both accurate and able to properly convey these elements. This is even more pronounced if you rely on OCR instead of doing manual transcripts (or at least being able to do them).

Best advice is to just learn Japanese, then you can use MTL as a sort of dictionary or phrasing helper, if needed.

 

[exlurker]

Grok it is uncensored

I also recommend grok. I have been using it for translating doubts in unteralterbach. It was invaluable for translating Annemarie's poetry.

 

[Darkshade20200]

Depending on how much money someone is willing to spend.
Claude Plattform (Haiku+Sonnet) are all 3 uncensored.

Or if you are willing to learn und have much time to spend, u could use a local AI instead like ggml.
But for a local, you need to look at the Frontend u use (if you wanna uncensored, look for one, without implemented filter)

 

[Matm95]

Of course LLMs don't beat human performance yet, and they probably wont for quite a while, but they come somewhat close.
Online option nobody has mentioned yet: DeepL
Local LLMs: Deepseek (Censored, but you can find abliterated versions) or the more lightweight Llama models with 8b parameters both perform well.
If youre translating things containing rape, underage characters, or other things that are morally questionable in fiction, then you probably have to look for a model without censoring.

 

[m72o5xvz81]

prompt injection can trick gpt4 into processing explicit material

You guys doing prompt injection on API services - aren't you a little concerned about flagging your account and being on a list?

I'm not even sure how you get GPT4/5 on API to do prompt injection without going direct to OpenAI. Is there a service that gives you ability to prompt inject without a string directly back to you?

 

[Matm95]

You guys doing prompt injection on API services - aren't you a little concerned about flagging your account and being on a list?

I'm not even sure how you get GPT4/5 on API to do prompt injection without going direct to OpenAI. Is there a service that gives you ability to prompt inject without a string directly back to you?

They probably either dont really give a shit about being traced back, or they are on a VPN combined with a dedicated/burner account