I don't know which RPG MAKER this is. It is neither RPG Maker VX Ace nor RPG Maker MZ nor RPG Maker MV.
All the text stored in \spt\ in simple txt files, only with different extension (spt). Just copy those files somewhere, change spt to txt and open with Notepad++ or something that can figure encoding automatically. What you need is the text in them that goes after msgt (not all files have them). Maybe there's some other spots you'll have to translate, but you can figure it yourself. You also can translate txt files with Sugoi (notice, that it can deal only with simple txt saved in urf-8), then make sure that non-msgt Japanese text stayed the same, and then change txt to spt.
Translator ++ probably can extract relevant text automatically, but if it can't, then you can just edit everything manually. Though it would be a pain in the ass without some sort of automation.
I have converted an SPT file to TXT to see if I can open it with Translator++, but nothing is displayed. I wanted to do an automatic translation to German and play it through to understand the context better and then correct it manually later. Anyway, I'll leave it, thanks for your help though.
Can you reupload this on other sites like gofiles or mega? Pixeldrain is blocked in my country
tysm
Entai2965
Member
- Jan 12, 2020
- 161
- 459
Can you reupload this on other sites like gofiles or mega? Pixeldrain is blocked in my country
You must be registered to see the links
it's happening , a new version is coming out on the 19.
You must be registered to see the links
Here's hoping its an actual updated model and not just the bloatware that barely works (For me, anyways)
Entai2965
Member
- Jan 12, 2020
- 161
- 459
That is unlikely to happen, ever. The developer already stated that there is not really a way to improve the current model further. It has run its course in terms of technological development and there would be, at best, only incremental updates possible now which makes it not really worth the GPU time to retrain the model. Hence, there is no reason for the developer to bother improving it anymore.
If you want better translations than what can Sugoi can produce now, that requires either adding dictionaries to hardcode specific translations like SLR Translator does which fixes a lot of quirks that Sugoi has, but leaves the subjects messed up of course like all NMTs, or just go ahead and use the superior technology of LLMs instead. The successor technology to the NMT technology used the Sugoi Toolkit is Large Language Models (LLMs), so if you want any practical improvements over Sugoi, then look into using AI translations instead.
I did a comparison between Sugoi, DeepL, Mixtral8x7b. The results were that Sugoi is better than LLMs without context, but with context, LLMs are better at the cost of significantly increased computational time and reduced automation. For the minimal computation time that it has, the Sugoi NMT model included Sugoi Offline Translator v4, included in Sugoi since Sugoi Toolkit v6, is the best quality realistically possible for any JPN->ENG NMT model.
Mixtral is pretty 'old' at this point. Try this one
You must be registered to see the links
Supposedly better and smaller than mixtral.
Figured, I just wanted to believe... I guess the only other thing I hope is that he improves his OCR (Also, probably unlikely)
Entai2965
Member
- Jan 12, 2020
- 161
- 459
You must be registered to see the links
According to the leaderboard that ranks JPN VN -> ENG TL as of today mid-July 2024:
| rank | model | Accuracy |
| 1 | openai/gpt-4o-2024-05-13 | 0.747988 |
| 2 | anthropic/claude-3.5-sonnet | 0.747447 |
| 4 | nvidia/nemotron-4-340b-instruct | 0.719268 |
| 5 | lmg-anon/vntl-gemma2-27b_q5_k_m | 0.703626 |
| 6 | qwen/qwen-2-72b-instruct | 0.696493 |
| 7 | openai/gpt-3.5-turbo-1106 | 0.694348 |
| 8 | lmg-anon/vntl-llama3-8b-q8_0 | 0.68871 |
| 9 | google/gemma-2-27b-it_Q5_K_M | 0.68277 |
| 11 | mistralai/mixtral-8x22b-instruct | 0.678332 |
| 12 | cohere/command-r-plus | 0.674124 |
| 18 | meta-llama/llama-3-70b-instruct_Q4_K_M | 0.658825 |
| 25 | meta-llama/llama-3-70b-instruct | 0.63304 |
| 28 | mistralai/mixtral-8x7b-instruct | 0.616399 |
| 31 | meta-llama/llama-3-8b-instruct_Q8_0 | 0.604868 |
| 32 | cohere/command-r | 0.601418 |
| - | Sugoi Translator | 0.6093 |
| - | Google Translate | 0.5395 |
| - | Naver Papago | 0.4560 |
| - | Alibaba Translate | 0.4089 |
The above results are not entirely believable. There is no way the quantized version of llama3-70b-instruct should perform better than the cloud version which makes me question the validity of the test.
In addition, the dataset used to train the models, and the test itself also included a lot of Kanji names. There is no way to correctly translate those without the person saying how their name should be said in the text. Since the vntl dataset includes a lot of those hardcoded mappings, if the test checks for them and considers them as part of the ranking, then the results are basically cheating and boosting the vntl models higher than they truthfully belong.
Still, it is an interesting leaderboard. If the results are taken at face value, vntl-gemma2-27b should be better than the llama3-8b version. And as I said earlier, and as my results showed, the difference between Sugoi and LLMs, especially without context is not very large. Sugoi holds up well given its limitations.
The developer has a
You must be registered to see the links
(scroll down of the link) if you want to follow along with the development of the toolkit. There was something mentioned about improving OCR, but I do not care about OCR much right now, so I did not read it very closely.
Last edited:
Quantization is a bit weird in a sense that it introduces more noise so to speak. It may have just tipped the scale in one particular run. Or it's just some quirk of llama.cpp since tokenizer may not be exactly one to one.
I didn't really check the methodology behind testing, but isn't evaluation set different from the training one?
I heard that the author behind the fine tunes doesn't really recommend gemma over llama 8b. I guess it makes sense because it's totally feasible to fit llama entirely in your gpu for blazing fast translation, while gemma is quite big and isn't that much better. (Plus google fudged something up with gemma2 release, no one knows what's up with it really.)
I was thinking of whipping up some sort of gui that would work with llama.cpp's server. I think it would be much cleaner solution than sugoi's 12 gigs of python env bloat. Llama 8b at q4_k_m (5gb) + server (70mb) + you could enable or disable gpu accelaration without hassle (or even use partial gpu accel.)
Shouldn't
You must be registered to see the links
already do this? I haven't tried anything yet, only read, so I don't know.Kobold is a chat frontend/launcher for llama.cpp. I meant something like sugoi's interface where you copy input into your clipboard and it sends it for translation.
Entai2965
Member
- Jan 12, 2020
- 161
- 459
Here are the release notes:
For the offline model:
"Sugoi Offline Model is now using CT2 package by default, replacing previous fairseq library. Accuracy is about the same while CPU processing speed is twice as fast (even more so when enabling GPU)."
You must be registered to see the links
For the offline model:
"Sugoi Offline Model is now using CT2 package by default, replacing previous fairseq library. Accuracy is about the same while CPU processing speed is twice as fast (even more so when enabling GPU)."
BoohooBitch
Member
- Oct 30, 2017
- 411
- 309
can we use sugio to translate offline text file? Can you please drop your thoughts on my TL question thread
https://f95zone.to/threads/best-opt...ere-offline-ai-modals-that-we-can-use.217624/
It can translate txt files, but it's better to copy lines into spreadsheet (Excel and the likes) and then use Sugoi with Translator++. With enough effort and ingenuity you can translate even those games that T++ can't parse on its own.
RazedToThrill
Member
- Feb 22, 2020
- 168
- 346
So it should be about the same speed as FuwaNovel's repackage of the tool into CTrans2?
Entai2965
Member
- Jan 12, 2020
- 161
- 459