- Jul 24, 2017
- 2,536
- 28,611
That's fair it's not for everybody, it's still a machine so maintaining consistency for everything is unlikely.
- Apr 7, 2021
- 430
- 912
You could just search "I obtained" and replace it with "Obtained" in every json file, like TL++ does, maybe? 
But keep going!
But keep going!
- Jul 24, 2017
- 2,536
- 28,611
Done and pushed to the patch, really that's all you have to report to me. I generally ignore stuff like that unless someone reports it since it doesn't bother me at all.You could just search "I obtained" and replace it with "Obtained" in every json file, like TL++ does, maybe?
But keep going!
- Apr 7, 2021
- 430
- 912
Understandable, some people just prefer some kind of consistency i guess and i thought maybe it would improve your so far good translation a little bit or appeal more to users like me that are just stucked to their consistency xD Thank you very much.
out of curiosity, doesn't openai have a no explicit policy? how well does the tool handle this?
- Jul 24, 2017
- 2,536
- 28,611
API Handles it fine with the right prompt. Most of the time doesn't complain translation wise.
question trying to play winter memories right now with the translations so far but how do i do it
- Jul 24, 2017
- 2,536
- 28,611
- Jul 11, 2017
- 931
- 708
- Jul 24, 2017
- 2,536
- 28,611
Haven't tried it. I have no idea what the cost of it is and Bard is generally considered worst than GPT4 so I just haven't bothered looking into it.
- Jun 26, 2018
- 130
- 339
Can you explain to me what part of this process is tied to a cost/fee, like I'm a 5yo?
I have no clue about any of this, but I find it very interesting!
I have no clue about any of this, but I find it very interesting!
- Jul 11, 2017
- 931
- 708
- Jul 24, 2017
- 2,536
- 28,611
Like how the estimator works? Sure.
It basically does exactly what the tool does when it's translating. It scans for a code, text, or whatever. Grabs it like normal, but instead of translating the text, it run it through a library (Created by OpenAI) that calculates the number of tokens in the text.
I take that, do the math according to the API cost, then total it all up. Then it spits out the total estimated cost and the cost of every file it scanned. Still not perfect but I find it's around 95% of the way there.
Yup, I've been using 4 some too as well. I think Bard uses Palm2 API for their devs and currently its in closed beta with rate limiting so I actually can't test it even if I wanted to. Once their pricing comes out if it's comparable to 3.5 I think I might take it for a spin.
Took a look at the open.ai plans for fun, is the "3.5 Boost Model" on frontpage post referring to "GPT-3.5 Turbo"? or is it something else completely?
- Jul 24, 2017
- 2,536
- 28,611
Might wanna take a look at Falcon 40B, or Falcon 7B... tbh I don't think 7B is up to the task, but 40B might be good enough, since it's open-source and you can run it locally (if you have the perfs for it), you won't need to deal with API filters if they ever become a problem, plus you won't be violating the liscence... which we are on a piracy website so I dunno how relevant that is
But the most important plus is that you could do some reinforcement training on the model to make it better at doing the translation, and can have more control over what the model outputs... there are plenty of already translated games, so maybe taking the original and translated version of those as training data could be a good idea to make it more suited to this work...
and it would be free.... well appart from the computer needed to run it, though I bet you can rent one online...
Edit: 40B needs 100GB of Ram... but 7B only needs 16... maybe with some reinforcement training it could work
Links:
Falcon 40B:
Falcon 7B:
But the most important plus is that you could do some reinforcement training on the model to make it better at doing the translation, and can have more control over what the model outputs... there are plenty of already translated games, so maybe taking the original and translated version of those as training data could be a good idea to make it more suited to this work...
and it would be free.... well appart from the computer needed to run it, though I bet you can rent one online...
Edit: 40B needs 100GB of Ram... but 7B only needs 16... maybe with some reinforcement training it could work
Links:
Falcon 40B:
You must be registered to see the links
Falcon 7B:
You must be registered to see the links
- Jul 24, 2017
- 2,536
- 28,611
I'm actually waiting for the community to come to a consensus on a model that is at least on par with GPT4. I've tried a lot of these models and all of them while good at natural English fail miserably at Japanese translation. Likely because they aren't trained on it. For example.
Falcon40b isn't trained on Japanese so I would not expect it to perform very well for translation. It takes time to check all these models individually so I prefer to just wait and see until something outstanding comes along. I'm not too versed in machine learning and honestly not super interested since its its own field. I'm just here to focus on the machine translation aspect and improving that as much as possible.
Ah yeah, hadn't noticed the bit about it not being trained on Japanese... Well hopefully in the next few months we'll get something, GPT3 is not that old and we've come far since it came out... so with the speed at which progress is being made, I wouldn't be surprised if within the year we'll get a good option