RPGM WIP Shisaye's Project List

[Shisaye]

Well, I can think of one that I'd play again in a heartbeat if you'd be willing to offer it more attention, and that would be Dark Beauty Reina:
RJ228148 Dark Beauty Reina Mermaid Island and the Saint of the Fort. 2.1
https://f95zone.to/threads/translat...saintess-v2-1-golden-cat.176659/post-11858954

Really any of the Golden Cat games are ones I'd play again...

As for a game that would specifically benefit from a better translation, I'd say RJ275785 Sheath Shoujo Aloof Girl:
https://f95zone.to/threads/translat...-parallel-world-rj275785.163074/post-11368438

When playing this one, I noticed the text lost...depth? Detail? Like the less obvious/blatant details weren't coming through in the translated text versus when I played it before using deepl. Not that deepl was perfect either, but something was definitely missing in the SLRMTL.

Just throwing my opinions out there...

Finished remaking Dark Beauty Reina with my current prototype. (Iteration 7 )
It took 3 hours, had over 80 errors all of which were automatically fixed by my system, the resulting translation did not trip a single SLR failsafe. Supposedly the translation is flawless. *Doubting intensifies*

I purposely did not do a single manual edit because it's about the capabilities of the raw engine.
https://f95zone.to/threads/dark-beauty-reina.57931/post-17056058


I will obviously do more testing, although I admit I didn't expect to actually get so many remake suggestions.

 

[Shisaye]

Is there anything you've translated that got an actual translation after the fact? I think that'd be the goal if you want to test the new system, shows how far the system has come and how close it is to "real" translation.

Please do not expect too much, I do not adjust the prompt to give DeepSeek context of the overall game's story, which means it's not actually going to be that big of a difference.
I've added functionality that a user "can" do that, but I want it to work for someone who has absolutely no clue what they're doing, so I need to test with universal prompts.

 

[Shisaye]

I would play to try and play both of these again if they had a better translation:

RJ315828 Forced NTR Dungeon -Female knight Rihanna v1.0
https://f95zone.to/threads/forced-ntr-dungeon-female-knight-rihanna.116126/#post-8070888

MOA RJ01145528 NTR Succubus v1.0
https://f95zone.to/threads/ntrsakyubasu-ntr-succubus.191754/#post-13734280

Here's a DSLRMTL for NTR Succubus.
Very short game though.
https://f95zone.to/threads/ntrsakyubasu-ntr-succubus.191754/post-17058073

 

[Shisaye]

Suzuka's Battle Record is a pretty small game and also has decent translation requests.
Village Girl Yuma and Barrier Master
Any Golden Cat game would also be appreciated.

I would play these for the 1st time

DSLRMTL for Village Girl Yuma and the Barrier Master:
https://f95zone.to/threads/yumato-rj01092914-landcut.196264/post-17060700

 

[Shisaye]

Well, I can think of one that I'd play again in a heartbeat if you'd be willing to offer it more attention, and that would be Dark Beauty Reina:
RJ228148 Dark Beauty Reina Mermaid Island and the Saint of the Fort. 2.1
https://f95zone.to/threads/translat...saintess-v2-1-golden-cat.176659/post-11858954

Really any of the Golden Cat games are ones I'd play again...

As for a game that would specifically benefit from a better translation, I'd say RJ275785 Sheath Shoujo Aloof Girl:
https://f95zone.to/threads/translat...-parallel-world-rj275785.163074/post-11368438

When playing this one, I noticed the text lost...depth? Detail? Like the less obvious/blatant details weren't coming through in the translated text versus when I played it before using deepl. Not that deepl was perfect either, but something was definitely missing in the SLRMTL.

Just throwing my opinions out there...

DSLRMTL for Dark Beauty Reina - Extra Story:
https://f95zone.to/threads/translat...saintess-v2-1-golden-cat.176659/post-17061885

I'm probably going to wait for some actual feedback now before making more, in case the translations are actually terrible/broken.

 

[Shisaye]

One thing I'm noticing is that DeepSeek seems to like to add "" around certain parts of the text even if the original text had nothing. Probably to differentiate between thought, noise, and someone saying something.
But not all the time, which makes it somewhat inconsistent.
I'm really not sure if I like that. It doesn't really fit the esthetic of the rest.
Of course Japanese does not have the punctuation of English, and so a translator adding stuff isn't necessarily wrong, but it just looks a bit weird.
For example when it comes to someone talking I've always preserved the Japanese 「『 』」boundaries to make speech bubbles instead.


What do you think? Should I make the system remove or replace "" placed by DeepSeek?
And if replace, replace with what?

(Specifically only talking about "" that are not present in the original Japanese text.)

Edit:
There's also cases in which it replaces the boundaries:

My normal SLR system would translate:
JapaneseText JapaneseText JapaneseText 『JapaneseTextThatisImportant』JapaneseText JapaneseText.
to:
EnglishText EnglishText EnglishText 『EnglishTextOfTheImportantThing』EnglishText EnglishText.

While DSLR would currently translate it to:
EnglishText EnglishText EnglishText "EnglishTextOfTheImportantThing" EnglishText EnglishText.


I'm not a native English speaker so I don't know if there's some emotional attachment to "", I personally think 『』is more pretty.

 

[c4145317]

Probably a question of who the audience is. Like, the people who want the honorifics maintained are going to be accustomed to the brackets. While the people who are not yet adjusted to all the sans, chans, and whatever are going to find another thing confusing them.

 

[fantasmic]

I prefer preserving 『』. It looks cleaner than "" and better separates out dialogue, which I think helps the overall readability in games which lack any nametags for the text boxes. You know, the kind which are just like

Ilulu
「Wait, I thought Ultra-Bacon was food, not an extremist philosophy...」

versus

Ilulu
"Wait, I thought Ultra-Bacon was food, not an extremist philosophy..."

It also looks cleaner if you have nested quotes and possessives. Something like
"I've heard you can buy alcohol at "Ramse's""
「 I've heard you can buy alcohol at 「 Ramse's 」」

A lot of this is a style choice, so if nothing else, it's a lot easier to use Find&Replace on [「『 』」] --> ["] than trying to manually replace each instance of ["] with the appropriate Japanese punctuation mark.

Ah, and to address the first question: I'd prefer to just remove any "" which Deepseek was trying to add for clarification. I feel most users are (or will become) familiar enough with Japanese and language in general to parse text themselves.

 

[Shisaye]

I prefer preserving 『』. It looks cleaner than "" and better separates out dialogue, which I think helps the overall readability in games which lack any nametags for the text boxes. You know, the kind which are just like

versus

It also looks cleaner if you have nested quotes and possessives. Something like
"I've heard you can buy alcohol at "Ramse's""
「 I've heard you can buy alcohol at 「 Ramse's 」」

A lot of this is a style choice, so if nothing else, it's a lot easier to use Find&Replace on [「『 』」] --> ["] than trying to manually replace each instance of ["] with the appropriate Japanese punctuation mark.

Ah, and to address the first question: I'd prefer to just remove any "" which Deepseek was trying to add for clarification. I feel most users are (or will become) familiar enough with Japanese and language in general to parse text themselves.

How about I tell deepseek to always preserve 「『 』」(as well as other styles from a long list) and not replace them with "".

And then this chain ( Text made more readable using deepseek ):

1. If the translation contains ":
   • Check if the original text has " or corner brackets (「『 』」).
     • If the original has no " or corner brackets: Remove all " from the translation.
     • If the original has corner brackets but no ":
       • Compare the number of corner brackets (「『 』」) in the original and translation.
         • If counts match: Remove all " from the translation.
         • If counts don’t match:
           • If the original has exactly 2 corner brackets (first and last characters):
             • Replace " with the correct corner bracket type (「 or 『) and remove extra ".
           • If the original has a different count:
             • Check if the " count in the translation is even.
               • If odd: Reject the translation and request a new one.
               • If even:
                 • If the translation has no corner brackets:
                   • If " count ≠ original’s corner bracket count: Reject and request a new translation.
                   • If counts match: Replace " with corner brackets in order of the original (first " → 「, second " → 『, etc.).
                 • If the translation already has corner brackets:
                   • If corner bracket count is odd: Reject and request a new translation.
                   • If even:
                     • Replace them, then check " count against the original’s corner bracket count.
                       • If they match: Replace " with corner brackets in order of the original (first " → 「, second " → 『, etc.).
                       • If not: Reject and request a new translation.
2. If none of the above applies: Accept the translation as-is.

Edit: And have it only request a new translation a maximum of 2 times before giving up, have a toggle for those retries in the options menu, and when it gives up have it accept the unedited first translation for this batch, but do not add it to the persistent cache.

 

[Shisaye]

If I were to do it like that (I'm still waiting for feedback) I could technically also make a new field in the options menu that allows you to put in a regex blacklist/whitelist.
In the sense that if the regex matches the translation given by deepseek in the blacklist case, it immediately rejects the translation and requests a new one.
If the regex matches the translation given by deepseek in the whitelist, it bypasses all other processing and accepts the translation.
And if it matches both the blacklist and the whitelist, it does nothing because otherwise it would be racist.

 

[Shisaye]

Hm, there's one really annoying issue I've noticed.
Since that's a general issue when translating Japanese to English (Japanese does not have pronouns), without a game specific prompt LLMs will still get pronouns wrong.
Now obviously you can just add "Ruka is a woman" to the prompt and it will try to keep text that looks like it could be her, with female pronouns, but if it already translated the text without it, it will have added the translations to the persistent cache and changing the prompt now does absolutely nothing because it would not even ask DeepSeek again, it would just load the old translations from cache.

But obviously you wouldn't want to delete your entire cache just for that.

Now you're gonna say "just make it so cache entries have a session ID so we can tell it to only delete cache from this session" but what if another game had those same sentences? You would still have no clue where in the cache they are...

Edit: The easiest solution would be to automatically disable the persistent cache when the "additional context" option is used, but that would make updating a game to a new version really dumb, because you couldn't just make it fill all unchanged cells from cache anymore and instead you would have to pay for the whole translation again. Also no more cache protection from crashes/power outages.

Edit2: Maybe I should just add a big fat note to the context box that you have to backup/move the cache file and if people don't it's not my fault. I mean if someone actually takes the effort to change the prompt, then they are probably capable of changing a cache file.

 

[Shisaye]

I've noticed some pretty serious issues with my old DSLR engine prototype. I've fixed almost all of them already in my current one, but those test translations are not representative of the format. Do no use them.

I will remake them with the new one soon, which will noticeably increase the translation quality. (Might still take a few days though.)

 

[Shisaye]

I would play to try and play both of these again if they had a better translation:
MOA RJ01145528 NTR Succubus v1.0
https://f95zone.to/threads/ntrsakyubasu-ntr-succubus.191754/#post-13734280

Village Girl Yuma and Barrier Master

I would play these for the 1st time

I've remade both translations with my current DSLR Prototype. (prot16 )

https://f95zone.to/threads/ntrsakyubasu-ntr-succubus.191754/post-17058073
(~30k characters of text)

https://f95zone.to/threads/yumato-rj01092914-landcut.196264/post-17060700
(~220k characters of text (But a third of it repeats))

Both combined took ~7h, but that includes checking and writing new code whenever I noticed something that could be improved.
With an average speed of 10 characters per second.
No "real" batch failures at all. (Just some server problems, that resolved themselves after trying again. (Response error status 500 followed by error status 429))
Not a single detected error after translation.
I only saw a single "obviously" bad translation while skipping through the maps, and it still conveyed the original meaning. (It just looked really ugly and was missing punctuation.)

Please be brutally honest about the translation quality. While I can make more improvements, they will likely only be incremental like adding another specific dictionary.
If this does not look "significantly" better than SLR then I will probably abandon the format since it costs money.
(I would still add general LLM support to SLR Translator though, because I've already made the engine anyway.)

I will make more test translations, but the now pretty sophisticated batch translation process takes super long.

Edit: Changed links to be rule compliant.

 

[Shisaye]

Working on it more I and having found, yet, another issue that makes translation quality worse than it has to be, I wanted to remake them, yet again, deleted my cache, yet again, and then it hit me that I think my approach to this testing is actually really stupid.

The only person that would use DSLR is someone who wants a better translation than SLR. To the point that they are even willing to pay deepseek for it.
If they already go that far for translation quality, why on earth would they not bother to put a custom context prompt to fix pronouns, characters names, etc.?

That would also mean that they would not want a persistent cache between translations or at least not one that is turned on by default, even if it means potentially paying for the same translation twice, because it would overwrite newly given instructions.
Because I mean we established that this user really cares about translation quality, because otherwise why on earth would they not just use normal SLR instead?

TLDR: I'm honestly not sure how to approach it, yet, but those 2 translations are still not representative of what I'm going for.
I think I'm going to take a few days to work out how exactly context prompting works and then I'll try again.

 

[Shisaye]

I've added limited automatic context now. Basically telling it it's last accepted translations within the same file.
Seems promising:
(At least Moose now accurately talks about her panties and not his own.)

 

[Shisaye]

Prototype 27 looks seriously promising now. It's now without excessive token usage keeping 5000 characters of context and paired with a custom context prompt it no longer fucks up gender and ownership at all.
Surely this time I wont find a major flaw again.

Surely it will just pass all testing and release me from this torment.

 

[wibuntr]

Prototype 27 looks seriously promising now. It's now without excessive token usage keeping 5000 characters of context and paired with a custom context prompt it no longer fucks up gender and ownership at all.
Surely this time I wont find a major flaw again.

Surely it will just pass all testing and release me from this torment.
View attachment 4910112

are we need to buy the token from deepseek api or its free??

 

[Shisaye]

are we need to buy the token from deepseek api or its free??

Until we can host big LLMs ourself we need to pay a provider.
They are quite varied in price.
But going by the absolute cheapest you could get away with translating a medium sized game for $0.40 worth of tokens (very slowly).

I'm trying to optimize it as much as I can, but if you want an actually good result completely free stuff wont work.
The most realistic "free" alternative right now would probably be a locally hosted Gemma3, but that would still need a monster GPU, hence I cannot test that.

I'm also planning to look into distilled models that might be feasible to host locally, but most of the time the smaller the model, the less it understands complex prompts.

Edit: From a shortish test with Prototpye 28, the average token usage each 2500 character batch is now 7000. That means a medium sized game (250k characters, but has repeats) is now around 600k tokens.
Using the fastest provider available on OpenRouter (Ignoring a certain super expensive one that isn't even really faster and has ridiculous prices) that would be around $0.46. Using the absolute cheapest that would be around $0.18.
(But that is assuming you are actually happy with the result and you aren't re-translating anything.)

 

[Osamu Dazai Sensei]

Until we can host big LLMs ourself we need to pay a provider.
They are quite varied in price.
But going by the absolute cheapest you could get away with translating a medium sized game for $0.40 worth of tokens (very slowly).

I'm trying to optimize it as much as I can, but if you want an actually good result completely free stuff wont work.
The most realistic "free" alternative right now would probably be a locally hosted Gemma3, but that would still need a monster GPU, hence I cannot test that.

I'm also planning to look into distilled models that might be feasible to host locally, but most of the time the smaller the model, the less it understands complex prompts.

Edit: From a shortish test with Prototpye 28, the average token usage each 2500 character batch is now 7000. That means a medium sized game (250k characters, but has repeats) is now around 600k tokens.
Using the fastest provider available on OpenRouter (Ignoring a certain super expensive one that isn't even really faster and has ridiculous prices) that would be around $0.46. Using the absolute cheapest that would be around $0.18.
(But that is assuming you are actually happy with the result and you aren't re-translating anything.)

Shisaye Can you also Translate this game while you're at it?

You must be registered to see the links

You must be registered to see the links

You must be registered to see the links

You must be registered to see the links

 

[Shisaye]

Shisaye Can you also Translate this game while you're at it?

You must be registered to see the links

You must be registered to see the links

You must be registered to see the links

You must be registered to see the links

I'm currently not taking random SLRMTL requests here, my hardware situation still has not improved and SugoiV4 is super slow. But SLR Translator is free and you can just do it yourself. Should you run into issues during the process then I will do what I can to help you fix it.


And just to have said it in general:
I hope I don't have to mention that I'm never going to take random DSLRMTL requests. I might consider doing those on a donation basis at some point, but I'm not really comfortable with that Idea and I doubt there's any demand either. (At least not with subscribestar being the only option.)
But I will release the DSLR Translation Engine with the next SLR Translator update, so people can play around with it themselves.