That seems a bit weird. I had some formatting issues when it came to commas, but was able to successfully translate a 14k line file at once.
- Nov 26, 2017
- 41
- 39
Here's the file. Perhaps because of rate limit on my account. About the screenshot i mean it should also output the rest of the rows (1000 in total) but the translated file only has 688 rows.
- Nov 26, 2017
- 41
- 39
Also, how do i stop it from adding (.) or (!) at the end of the translated text. i tried "You MUST not add any Period(.) or Exclamation Point(!) at the end of the translated text." and others similar but it didn't work.
- Nov 26, 2017
- 41
- 39
- Jul 24, 2017
- 2,531
- 28,587
Getting rid of the . or ! will require you to manually edit it or mess with the code. If you aren't code savy, its best to just manually remove that stuff. ChatGPT isn't good at following those kind of directions so he won't remove them. I would recommend using the python strip() function right before it sets data.
For the failed line, that's because ChatGPT fails to come up with a translation due to it being a long sound effect. It happens rarely but requires you to just manually edit that line to fix it.
Yeah from what I can see the csv functionality needs further testing since when ChatGPT returns some nonsense it add extra lines for some reason. I'll have to look into it more.
Might be your text editor, looking into excel the outputted number of lines is the exact same as the input. Even in your screenshot the line you drew points to the same line #.
I also see that you are translating an RPGMaker game that was converted to csv first? Any reason for that?
For the failed line, that's because ChatGPT fails to come up with a translation due to it being a long sound effect. It happens rarely but requires you to just manually edit that line to fix it.
Might be your text editor, looking into excel the outputted number of lines is the exact same as the input. Even in your screenshot the line you drew points to the same line #.
I also see that you are translating an RPGMaker game that was converted to csv first? Any reason for that?
Last edited:
- Nov 26, 2017
- 41
- 39
Because rows like this, it'll return incorrectly format:
And some name like "\N[4]" it just return "[Name]".
The GPT built-in in Translator++ is better with format but Batch Translation suck ass so i use this as alternative to the Batch Translation. Aslo because i dont have GPT Plus so it'll easily reach limit when translate big json file.
Last edited:
- Jul 24, 2017
- 2,531
- 28,587
It skips anything already translated so you can translate in parts.
- Dec 25, 2019
- 1,763
- 3,147
Hey dazedanon good news here 
Just recently was released
Said that, this model is still censored so will need the usual tricks but it shouldn't be long anyway to get a full uncensored.
Just recently was released
You must be registered to see the links
that work pretty much as Chat GPT with no exception, is a 4bit quantized model but it will reflect also the same performance as the original meaning that you can host it locally, now.. to host a 65b model even in 4bit you still need a large sum of vram (around 45gb +) but you can rent a server with these specs in a much cheaper way than to use the official api (eg Runpod will ask you 0.79$ /hr for 48gb), ofc if you not run it only for test. (Are available also the 33b, 13b, and 7b models and these can actually run in a consumer pc but ofc will effect the translation accuracy)Said that, this model is still censored so will need the usual tricks but it shouldn't be long anyway to get a full uncensored.
Last edited:
- Jul 24, 2017
- 2,531
- 28,587
Well all these models have claimed the same thing and its never true... even then the main problem is that none of them are actually trained in any Japanese so their translations usually suck. I'd need some proof on its translation capabilities before I'd actually try it.Hey dazedanon good news here
Just recently was releasedYou must be registered to see the linksthat work pretty much as Chat GPT with no exception, is a 4bit quantized model but it will reflect also the same performance as the original meaning that you can host it locally, now.. to host a 65b model even in 4bit you still need a large sum of vram (around 45gb +) but you can rent a server with these specs in a much cheaper way than to use the official api (eg Runpod will ask you 0.79$ /hr for 48gb), ofc if you not run it only for test.
Said that, this model is still censored so will need the usual tricks but it shouldn't be long anyway to get a full uncensored.
- Dec 25, 2019
- 1,763
- 3,147
Test in various languages have proven true, that's the only reason why i'm talking about.
- Jul 24, 2017
- 2,531
- 28,587
- Dec 25, 2019
- 1,763
- 3,147
Sorry, I have no video to show, try it yourself to see the results (sorry if happen to be presumptuous from me).
- Jul 24, 2017
- 2,531
- 28,587
Well I would if I could actually run these locally but since they require renting a server I'll pass for now.
- Dec 25, 2019
- 1,763
- 3,147
- Jul 24, 2017
- 2,531
- 28,587
Ehhh, it's not that great.
Edit:
Actually I've tested it with a few more and its not that bad for some things. I wonder how it would do with harder phrases. It's not enough though. I'm more interest in models that are specifically trained on Japanese.
Last edited:
- Dec 25, 2019
- 1,763
- 3,147
Keep in mind that you are actually using only the 13b model so is not a clear comparation, not by chance I've spoken about the 65b (still even the 13b so far is superior to any other LLaMa model with the same capacity).
Last edited:
- Jul 24, 2017
- 2,531
- 28,587
Yeah but I can't run the 65b without renting a server and there is no guarantee the 65b is any better at translation than 13b. The only way to grab my interest on these is if its specifically trained for Japanese like with NeoXJapanese.
Plus even if I were to find one I'd have to think about exactly how I would implement it considering my software makes specific use of an API. I'm not exactly a great programmer so anything super complicated but would probably take me a long time to understand.
- Dec 25, 2019
- 1,763
- 3,147
Tried
You must be registered to see the links
and it does a terrible job, not sure why would you recommend it, here anyway the
You must be registered to see the links
, the 65b will come in few days, right now the 33b still lack in dictionary so it wouldn't be a right comparation but you should notice already some improvements from the 13b.