Tool RPGM SLR Translator - Offline JP to EN Translation for RPG Maker VX, VX Ace, MV, and MZ

[Shisaye]

I've released SLR Translator v1.136.

Fixed a bunch of (cosmetic) bugs of the new batch translation and improved the way full batch interacts with the endpoint, as well as error handling, automatic correction requests, etc.

During some random tests with gemma3 12B performance was now significantly better than with v1.135.

One of the new "fixes" is that it will no longer request a correction if a llm has 'ー' in the response, because apparently models like gemma3 are too stupid to recognize it as a Japanese character.
Even if you give it an assistant prompt with the wrong response labeled wrong and put a big fat never use Japanese characters in a critical instructions user prompt, gemma3 will just give you the exact same wrong response again until the system gives up after 12 attempts and the whole batch fails.

 

[RenderedFatality]

I've released SLR Translator v1.136.

Fixed a bunch of (cosmetic) bugs of the new batch translation and improved the way full batch interacts with the endpoint, as well as error handling, automatic correction requests, etc.

During some random tests with gemma3 12B performance was now significantly better than with v1.135.

One of the new "fixes" is that it will no longer request a correction if a llm has 'ー' in the response, because apparently models like gemma3 are too stupid to recognize it as a Japanese character.
Even if you give it an assistant prompt with the wrong response labeled wrong and put a big fat never use Japanese characters in a critical instructions user prompt, gemma3 will just give you the exact same wrong response again until the system gives up after 12 attempts and the whole batch fails.

Quick question, do you have an idea of what might be limiting concurrent batches to 6 queries in parallel no matter what? tried raising the batch size limit, then the batch length, it seems to just not want to go past that

Tried probing the files to see what might be the issue, and I've no clue.

 

[Shisaye]

Id like some suggestions for local models to test, because I still haven't found a single one that I would consider using.
I don't really want to use one myself because I'm perfectly happy with DeepSeek, but I'd like to actually be able to give people some kind of "free" example to be able to fully test the DSLR system.

The ones I tried so far are:

mistral-small 3.2 (Q4_K_M)
Failed to keep linecount for script sections even after 12 attempts ignores correction instructions.

mistralai/magistral-small (Q4_K_M)
Failed to keep linecount for script sections even after 12 attempts ignores correction instructions.

mistralai/magistral-small-2506 (Q4_K_S)
Failed to keep linecount for script sections even after 12 attempts ignores correction instructions.

DeepSeek-R1-0528-qwen-3-8B (Q4_K_M)
Crashed in full batch mode.
I aborted the test in single translation mode because it would have taken too long.

DeepSeek-R1-Distill-Llama-8B (Q4_K_S)
Crashed in full batch mode.
Ignores instructions not to include the original text in the response.

meta-llama-3.1-8b-instruct (Q4_K_M)
Ignores instructions not to include the original text in the response.
Eventually shit the bed and was just repeating the same broken response
Does work a lot better in individual translation mode, but the translation quality is not good.

vntl-llama-8B (Q4_K_M)
Does not understand complex prompts, basically just SugoiV4 in slower.

Gemma 2 9B (Q3_K_L)
Started hallucinating entries and refused to correct itself.

Gemma 3 4B (Q3_K_L)
Started hallucinating entries and refused to correct itself.

Gemma 3 12B (Q4_K_M)
Much better than 4B, but eventually still hallucinated and shit the bed.

Gemma 3 27B (Q4_0)
First model to actually finish the test, but very slow and the translation quality wasn't super great so it hardly seems worth using instead of the literally x1000 faster SugoiV4.

Edit: Also in case that's not obvious if I throw the exact same test at deepseek-chat-v3, it doesn't even need retries. No hiccups at all.

 

[Shisaye]

Quick question, do you have an idea of what might be limiting concurrent batches to 6 queries in parallel no matter what? tried raising the batch size limit, then the batch length, it seems to just not want to go past that

Tried probing the files to see what might be the issue, and I've no clue.

Might honestly be a Node-js limitation, because I have not given it any cap in the code.
As you've probably already seen it's handled by the "RedBatchTranslator" class in the "transredbatch.js" found in www>addons>SLRbatch.

 

[RenderedFatality]

Qwen 4B no think seems to do about as well as i would expect, need a really tiny batch size, like 100 characters. the issue with smaller models is that it will do a
View attachment 2025-06-28 06-42-02.mp4
romanized translation rather than a english translation, and there is no good way to fix that

 

[Shisaye]

Qwen 4B no think seems to do about as well as i would expect, need a really tiny batch size, like 100 characters. the issue with smaller models is that it will do a
romanized translation rather than a english translation, and there is no good way to fix that

I've included a toggle in the options menu that can make it no longer reject responses based on leftover Japanese, but I feel like at that point you might as well just use normal SLR because while that doesn't give a great translation at least it never includes moonrunes.

 

[Shisaye]

Tried more:

phi-3.5-mini-instruct_uncensored Q8_0
darkidol-llama-3.1-8b-instruct-1.2-uncensored
L3.2-rogue-creative-instinct-uncensored-abliterated-7b Q8.0
Llama-3.2-8X3B-MOE-Dark-Champion-Instruct-abliterated-18.4B Q4_K_S
Llama-3.2-8X3B-MOE-V2-Dark-Champion-Instruct-abliterated-21B Q4_K_S
SOLAR-10.7B-Instruct-v1.0-uncensored Q8_0

All failed the stress test.

Best performer was SOLAR-10.7B-Instruct-v1.0-uncensored Q8_0, but only in single translation mode, which is super slow for a model that size.

 

[Samjaza]

anyone tried with the Sugoi 14B/32B yet?

 

[Shisaye]

anyone tried with the Sugoi 14B/32B yet?

I want to be clear that this is only about the first versions released, and does not necessarily mean it will stay this way:

Currently the Sugoi LLMs do not really understand instructions, they are almost exclusively "Japanese Text in" > "English Text out".
They also do not preserve things like text effects, placeholders, etc.
As a result they can not be used with any of the advanced features of DSLR and instead can only be used in a similar fashion to SugoiV4, but also are not really suitable as a replacement for SugoiV4, because they require substantial hardware to run, while SugoiV4 does not.

 

[RenderedFatality]

Tried more:

phi-3.5-mini-instruct_uncensored Q8_0
darkidol-llama-3.1-8b-instruct-1.2-uncensored
L3.2-rogue-creative-instinct-uncensored-abliterated-7b Q8.0
Llama-3.2-8X3B-MOE-Dark-Champion-Instruct-abliterated-18.4B Q4_K_S
Llama-3.2-8X3B-MOE-V2-Dark-Champion-Instruct-abliterated-21B Q4_K_S
SOLAR-10.7B-Instruct-v1.0-uncensored Q8_0

All failed the stress test.

Best performer was SOLAR-10.7B-Instruct-v1.0-uncensored Q8_0, but only in single translation mode, which is super slow for a model that size.

qwen3-30b-a3b at iq2_xxs works, does fairly well with 400 character batch sizes as well.

i did a small test with a 2k batch size on a small file, it did it once, so i think its probably your best bet, but it definitely isnt consistent doing 2k characters.

and it gets to about 19 t/s on cpu only inference on a 9800x3d,

22 t/s when offloading 12/48 layers to the gpu

26 t/s when offloading 24/48 layers to the gpu

32t/s when offloading 36/48 layers to the gpu

160 t/s when offloading 48/48 layers to the gpu

 

[Samjaza]

I want to be clear that this is only about the first versions released, and does not necessarily mean it will stay this way:

Currently the Sugoi LLMs do not really understand instructions, they are almost exclusively "Japanese Text in" > "English Text out".
They also do not preserve things like text effects, placeholders, etc.
As a result they can not be used with any of the advanced features of DSLR and instead can only be used in a similar fashion to SugoiV4, but also are not really suitable as a replacement for SugoiV4, because they require substantial hardware to run, while SugoiV4 does not.

Is it worth trying to get it working or just use something like "SOLAR-10.7B-Instruct-v1.0-uncensored Q8_0" .

 

[Shisaye]

Is it worth trying to get it working or just use something like "SOLAR-10.7B-Instruct-v1.0-uncensored Q8_0" .

The reason I'm trying out so much random stuff is because I still haven't found a single LLM, that can be locally run on "average" hardware that I would actually recommend.
I'm just trying to find a single one I could include as an example in the tool, but no luck.
My actually best result so far was Gemma3-12B-Q8_0, but eventually it hallucinated and went into some kind of weird loop of just repeating the same thing until retry requests were exhausted.
Gemma3 27B finished the test, but that might have been a fluke, and it was so dreadfully slow on my machine that I can't possibly recommend it.
Ultimately the only thing I can recommend with confidence right now is actually paying for DeepSeek. It hasn't failed me, yet.

qwen3-30b-a3b at iq2_xxs works, does fairly well with 400 character batch sizes as well.

i did a small test with a 2k batch size on a small file, it did it once, so i think its probably your best bet, but it definitely isnt consistent doing 2k characters.

and it gets to about 19 t/s on cpu only inference on a 9800x3d,

22 t/s when offloading 12/48 layers to the gpu

26 t/s when offloading 24/48 layers to the gpu

32t/s when offloading 36/48 layers to the gpu

160 t/s when offloading 48/48 layers to the gpu

I'll try that one.

I have tried Qwen3 14B Q8_0, but that one failed the test. It did not keep the dividers in full batch mode (500 character batches), did not keep linecount for scripts in individual translation mode, and refused to correct itself.

Edit: Err... where do I get that one?
Edit2: Nvm found it. I've never used a Q2_XXS quant before.

 

[RenderedFatality]

The reason I'm trying out so much random stuff is because I still haven't found a single LLM, that can be locally run on "average" hardware that I would actually recommend.
I'm just trying to find a single one I could include as an example in the tool, but no luck.
My actually best result so far was Gemma3-12B-Q8_0, but eventually it hallucinated and went into some kind of weird loop of just repeating the same thing until retry requests were exhausted.
Gemma3 27B finished the test, but that might have been a fluke, and it was so dreadfully slow on my machine that I can't possibly recommend it.
Ultimately the only thing I can recommend with confidence right now is actually paying for DeepSeek. It hasn't failed me, yet.

I'll try that one.

I have tried Qwen3 14B Q8_0, but that one failed the test. It did not keep the dividers in full batch mode (500 character batches), did not keep linecount for scripts in individual translation mode, and refused to correct itself.

Edit: Err... where do I get that one?
Edit2: Nvm found it. I've never used a Q2_XXS quant before.

its about as close to quant brain damage as you can get actually.

 

[RenderedFatality]

And i just tested qwen 14b q6k, are you making to sure turn off some of the lmstudio defaults? like the repeat penalty and the min-k? the qwen series doesnt like either too much

and honestly even the a3b ate it up at 500 char batch
View attachment 2025-06-30 07-17-27.mp4
View attachment 2025-06-30 07-11-59.mp4

 

[Shisaye]

And i just tested qwen 14b q6k, are you making to sure turn off some of the lmstudio defaults? like the repeat penalty and the min-k? the qwen series doesnt like either too much

and honestly even the a3b ate it up at 500 char batch
View attachment 4995873
View attachment 4995849

I can try again with that, I actually left it on default settings since I figured that's what most people would do, but I don't just feed it plain text. It needs to include text effects like color changing codes. (Text \C[12]Word\C[0] Text) or general placeholders like @1 @2 etc.
I want the model to be able to translate an entire somewhat complicated game in one go.

 

[RenderedFatality]

I can try again with that, I actually left it on default settings since I figured that's what most people would do, but I don't just feed it plain text. It needs to include text effects like color changing codes. (Text \C[12]Word\C[0] Text) or general placeholders like @1 @2 etc.
I want the model to be able to translate an entire somewhat complicated game in one go.

well, that's fair. Both of these seemed to handle codes fine at 500 batch after configuring them.

 

[Shisaye]

well, that's fair. Both of these seemed to handle codes fine at 500 batch after configuring them.

View attachment 4995931

Unless I fucked up those ones should never actually make it to the LLM, they should be escaped and later re-inserted.
If they are actually sent then I have much bigger problems than the model choice.

Edit:In fact that does already look like I fucked up.... should those even be translated? What plugin is that?
Edit2: Or rather what game is that (DLsite code)? I don't think this is working correctly...

Edit3: I'm really not properly escaping \E[] commands it seems. I don't know what those are supposed to do, but judging from the Japanese I would assume it's either showing pictures or playing sounds. Not something that can be translated. I will fix that in the next version. Thanks for showing me lol.

 

[Shisaye]

well, that's fair. Both of these seemed to handle codes fine at 500 batch after configuring them.

View attachment 4995931

Fixed it not properly escaping \E[] command content in v1.138.
Edit:
I found out my issue with qwen.
Apparently my code for removing the think block can fail somehow and leave part of it in.
It works most of the time but then suddenly fails:

with parts of the block left in:

 

[Shisaye]

I've released v1.139.
Fixed a really stupid bug that was sometimes causing a "think" block of a reasoning model to only be partially removed causing the response to be rejected.
That also means I have to retest all reasoning models I've tried so far, because the bad performance might have been my fault.

 

[Shisaye]

I finally got a winner.

It passed the entire stress test.
It needed a lot of retries for complicated script stuff, but ultimately always corrected itself.
Speed wasn't great, but certainly better than Gemma3-27B.
Translation quality of plain text is very underwhelming, but if you throw complicated stuff at it, and/or give it a game specific context prompt, it does actually beat SLR.

So if you absolutely do not want to pay for DeepSeek (which is still better obviously), or you absolutely do not want to connect your Windows machine to the internet (which I can fully understand), and you want to test the capabilities of the DSLR engine, then I can now recommend the reasoning model:

qwen3-30b-a3b-abliterated-erotic-i1-IQ3_XSS

(Or a better quant, IQ3-XSS is just what I used for the test.)

I've added the DSLR configuration I've used for that model in the stress test to the documentation in v1.140.
I've also fixed a bunch of small bugs and made the reasoning block removal system much more robust.