RPGM Completed WIP Quarta Knight Emerald [1.15][Kudarizaka Guardrail]

[yts1083]

+2 weeks.

Making good progress on Stage 3 though. I got through most of the combat erotic content already, and am now in the defeat scenes, so hoping to get it done sometime in January. Found an odd thing in the defeat scenes though. There appears to be two scenes for the Datajills, but I am not sure what actually triggers the second. Tried with it being the second scene or after other defeats, no trigger in the Reminiscence Room, with high degree of corruption. All duds. A mystery!

I've been using Google Translate to double check things, and it's surprisingly good. I'll be doing some experiments with that as well for Stage 4 and 5 to see if the translation work goes faster with that being the base instead of DeepL.

keep up the good work! I wonder why Kagura never picked it up as they did the other one

 

[candyate]

I bursted out laughing at the "ordinary train enjoyer" translation, well done

 

[MuninKai]

TRANSLATION UPDATE

You must be registered to see the links

is now available.

This is Stage 3 complete, and Stage 4 and 5 redone in Gemini MTL. Did some spot checks and the defeat scenes for Stages 4 and 5 are readable. Gemini is also way better and requires much less correction. I can do a defeat scene in about 45 minutes with Gemini compared to 2 or 3 hours with DeepL, so I am hoping that will accelerate the translation pace.

Spoiler: CHANGELOG

You don't have permission to view the spoiler content. Log in or register now.

 

[lilpixxle]

TRANSLATION UPDATE

You must be registered to see the links

is now available.

This is Stage 3 complete, and Stage 4 and 5 redone in Gemini MTL. Did some spot checks and the defeat scenes for Stages 4 and 5 are readable. Gemini is also way better and requires much less correction. I can do a defeat scene in about 45 minutes with Gemini compared to 2 or 3 hours with DeepL, so I am hoping that will accelerate the translation pace.

Spoiler: CHANGELOG

You don't have permission to view the spoiler content. Log in or register now.

awsome, thanks for all your hard work mate

 

[bazelgeuse]

Sounds basically done

 

[sportsfan30]

TRANSLATION UPDATE

You must be registered to see the links

is now available.

This is Stage 3 complete, and Stage 4 and 5 redone in Gemini MTL. Did some spot checks and the defeat scenes for Stages 4 and 5 are readable. Gemini is also way better and requires much less correction. I can do a defeat scene in about 45 minutes with Gemini compared to 2 or 3 hours with DeepL, so I am hoping that will accelerate the translation pace.

Spoiler: CHANGELOG

You don't have permission to view the spoiler content. Log in or register now.

How did you set up Gemini to allow nsfw content and are you able to bypass the token limits? Been debating on using AI tools for an initial pass on future projects.

 

[MuninKai]

How did you set up Gemini to allow nsfw content and are you able to bypass the token limits? Been debating on using AI tools for an initial pass on future projects.

The NSFW content is quite easy. I am doing it through the API, so I just pass it the following as part of the generate_content() function in Python. This is in the documentation as well.

safety_settings={HarmCategory.HARM_CATEGORY_SEXUALLY_EXPLICIT: HarmBlockThreshold.BLOCK_NONE, HarmCategory.HARM_CATEGORY_DANGEROUS_CONTENT: HarmBlockThreshold.BLOCK_NONE}

For the token limits, Google offers a 300$ credit when you first sign up that expires after 3 months. That being said, it's so cheap I won't be mad when that expires. I completely redid Goblin Burrow: Fable, I think the largest game I've ever tackled in terms of text, for 1.80$. The quality improvement over DeepL, and the time savings when doing edits, are worth it. EDIT: Compared to estimates from ChatGPT using Dazed's tool, this is manageable for me without crowdsourcing money to run it. The output is probably worse, but hey, get what you pay for.

If that is still too much, Google gates Gemini on the free plan by requests per day far more strictly than by token use per day or per request through the API. So I ran some experiments with creating batches of lines to translate. If I was really aggressive with the aggregation, I could do 50,000-75,000 (Google says you have 1,500 requests per day, but it cut me off around 1,100) lines per day. I found going above 50 lines per request resulted in a lot of garbage, so that's what I set it at. Monitoring my usage, I capped out on daily requests while using about 4% of the daily free tokens.

You do lose anonymity. Technically all my translation work is now visible to Google and tied to my real name and credit card, which is not ideal, but probably not going to be an issue.

The biggest issue I've identified is that Gemini is... a fucking idiot basically. I was extremely skeptical of LLMs before using one, and find the hype to be entirely unearned after using it. It follows instructions at best some of the time. Even with explicit instructions about, for example, how to translate specific names, it will ignore them about 50-75% of the time. There was a lot of prompt engineering as I was trying to wrestle with this very powerful, stupid, baby brained computer, and I will probably continue iterating my prompts for a while. It does produce better English than DeepL, and is a little more flexible than Google Translate, but it's still stupid and is especially stupid at scale when a human can't babysit its every output.

 

[sportsfan30]

The NSFW content is quite easy. I am doing it through the API, so I just pass it the following as part of the generate_content() function in Python. This is in the documentation as well.

safety_settings={HarmCategory.HARM_CATEGORY_SEXUALLY_EXPLICIT: HarmBlockThreshold.BLOCK_NONE, HarmCategory.HARM_CATEGORY_DANGEROUS_CONTENT: HarmBlockThreshold.BLOCK_NONE}

For the token limits, Google offers a 300$ credit when you first sign up that expires after 3 months. That being said, it's so cheap I won't be mad when that expires. I completely redid Goblin Burrow: Fable, I think the largest game I've ever tackled in terms of text, for 1.80$. The quality improvement over DeepL, and the time savings when doing edits, are worth it. EDIT: Compared to estimates from ChatGPT using Dazed's tool, this is manageable for me without crowdsourcing money to run it. The output is probably worse, but hey, get what you pay for.

If that is still too much, Google gates Gemini on the free plan by requests per day far more strictly than by token use per day or per request through the API. So I ran some experiments with creating batches of lines to translate. If I was really aggressive with the aggregation, I could do 50,000-75,000 (Google says you have 1,500 requests per day, but it cut me off around 1,100) lines per day. I found going above 50 lines per request resulted in a lot of garbage, so that's what I set it at. Monitoring my usage, I capped out on daily requests while using about 4% of the daily free tokens.

You do lose anonymity. Technically all my translation work is now visible to Google and tied to my real name and credit card, which is not ideal, but probably not going to be an issue.

The biggest issue I've identified is that Gemini is... a fucking idiot basically. I was extremely skeptical of LLMs before using one, and find the hype to be entirely unearned after using it. It follows instructions at best some of the time. Even with explicit instructions about, for example, how to translate specific names, it will ignore them about 50-75% of the time. There was a lot of prompt engineering as I was trying to wrestle with this very powerful, stupid, baby brained computer, and I will probably continue iterating my prompts for a while. It does produce better English than DeepL, and is a little more flexible than Google Translate, but it's still stupid and is especially stupid at scale when a human can't babysit its every output.

Thanks for the info and tips. I swapped to Linux awhile back and haven't figured out how to get Sugoi to work. I'm still looking at other options to supplement DeepL and Google Translate and wasn't sure how useful these glorified chat bots would be.

 

[MuninKai]

Thanks for the info and tips. I swapped to Linux awhile back and haven't figured out how to get Sugoi to work. I'm still looking at other options to supplement DeepL and Google Translate and wasn't sure how useful these glorified chat bots would be.

If you have a good process for Google Translate, I'd say it's probably not worth it to change over. In January I was looking at swapping from DeepL to either of Google Translate or Gemini, so I was designing a new process either way. When Gemini outputs garbage, which is rare, I plug it into Google Translate. If you do make the switch to Gemini, let me know and I can send you the prompts and structured output format I'm using.

 

[sportsfan30]

If you have a good process for Google Translate, I'd say it's probably not worth it to change over. In January I was looking at swapping from DeepL to either of Google Translate or Gemini, so I was designing a new process either way. When Gemini outputs garbage, which is rare, I plug it into Google Translate. If you do make the switch to Gemini, let me know and I can send you the prompts and structured output format I'm using.

Sure, I would like to take a look and see if it's worth moving over to it.