About translation with DeepSeek

[colemar]

I have tried various tricks with ChatGPT and Gemini web sites, but I failed to make them translate one entire subtitle file.
You cannot upload a file, the translation stops, etc. etc. There is always some issue.
I have a one month trial for Gemini and only the free ChatGPT version.

DeepSeek is often busy, but when available it seems to do a better job (a bit slow tough):

[niksedk]

I've tested DeepSeek locally via Ollama. The 4GB models is not very good. The 9GB models seems a lot better, but it's slow.

[darnn]

For what it's worth, this is what I do with any of these: In SE, I export the text of the subtitles (File->Export->Plain text), setting Format text to none. Then I feed the text in chunks to the LLM, whatever it may be (in my case usually ChatGPT), with the following prompt:

The following is text from subtitles. Translate it into text for English subtitles. Don't use text from previous prompts or say "here's the translation". Provide just the translation itself. Keep the formatting of the text the same, that is to say, the line breaks and empty lines between subtitles. The output should have the same number of subtitles as the input. If a line starts mid-sentence, don't capitalize the start of it. Otherwise capitalize normally:

The capitalization stuff is less relevant in languages where that's not a factor, but you get the idea.

While translating into English this is usually not necessary, but you can also do something like this:

I want you to translate some text into (language). Before that, I'm going to give you some English text, and then the way I translated it into (language) myself. I want you to translate in the same way I did. I will now provide the sample of English text that I've translated.
And then give it the original and the translation as stated. This doesn't make a huge difference, but it does help a bit. However, at least with ChatGPT, it's best to start a new session and do this every time you're translating something, because the more you use a single session the worse it gets, for some reason.

This is all to say that if what you want to use is DeepSeek, if you do the text export thing, you don't have to do it locally, you can just do it here:
https://chat.deepseek.com/
And you might also want to check out c4ai-aya-expanse-32b here:
https://dashboard.cohere.com/playground/chat

They're both worse than ChatGPT for my purposes, but not by that much.

[colemar]

@darnn Can you translate one file of about 900 subtitles in one go?

[darnn]

Not in one go. I give it chunks, usually of ~450 words. Finding the right amount requires some trial and error - you shouldn't do the maximum number of characters you're allowed to do in a single input. For me after about 450 words the quality of the translation starts to suffer, and sometimes it starts actually condensing several sentences into one and such. But it's still more effective than giving it the timecodes as well, since that's not really necessary, and (depending on how long your 900 subtitles are) it shouldn't take more than five minutes or so of sending the entire file in chunks like this. At any rate, it's the fastest way I've found of doing this, and the one that produces the best results.

Also, I guess I didn't mention the second half of this because to me it goes without saying, but once you have the full translation, you do the reverse, use Import plain text so that you have the translated text as a subtitle file, copy the contents of that file, and then use Column paste to paste just the text of it into an empty translation file in SE, so that you get the translated text with the timecodes of the original. I can elaborate on this part if needed.

[Name2889]

Could someone please explain how to set up offline AI for subtitle translation? I have never done it before, so if possible, please start from the basics. I need daily translations from Hindi to Russian. My mother watches an Indian drama, and I used to do it via Subtitle Edit and Google Translate V1. Now, it has become impossible, as more than half of the subtitles remain untranslated. I have been using MyMemory Translate for several days, but now it says that there are too many requests and has stopped working.
Thank you.