Which LLM model is best for translations?

[Dydhzo]

Hello,

I would like to know if you have any LLM models to recommend. I'm looking for one because I haven't found a model that works really well yet. I use a 3090 for my translations into French.

I've tried several models, including:

• Qwen2.5:32b:

  • Issue: Sometimes, it doesn't translate certain segments, for example, from Korean (KR) to French (FR).
    Example:
    Original:

    62
    00:03:23,953 --> 00:03:26,039
    [태영] 무고한 사람을
    위력으로 결박해 끌고 오는 등
    

    Translation:

    62
    00:03:23,953 --> 00:03:26,039
    [태영] Enchaîner et traîner des personnes innocentes par la force,
    

  • The model also has recurring issues, such as confusion between masculine and feminine.
    Despite this, I find Qwen2.5:32b overall quite good, but these problems (especially untranslated words) happen frequently.

• aya-expanse:32b:
  I strongly advise against using this model. It's really bad, complete nonsense.

I've also tried other models, but they were even worse.

So, I'm looking for a model that works well. I don't expect perfection, but at least translations that are complete and respect gender (masculine/feminine) properly.

[Neoony]

Hello,

I would like to know if you have any LLM models to recommend. I'm looking for one because I haven't found a model that works really well yet. I use a 3090 for my translations into French.

I've tried several models, including:

• Qwen2.5:32b:

  • Issue: Sometimes, it doesn't translate certain segments, for example, from Korean (KR) to French (FR).
    Example:
    Original:

    62
    00:03:23,953 --> 00:03:26,039
    [태영] 무고한 사람을
    위력으로 결박해 끌고 오는 등
    

    Translation:

    62
    00:03:23,953 --> 00:03:26,039
    [태영] Enchaîner et traîner des personnes innocentes par la force,
    

  • The model also has recurring issues, such as confusion between masculine and feminine.
    Despite this, I find Qwen2.5:32b overall quite good, but these problems (especially untranslated words) happen frequently.

• aya-expanse:32b:
  I strongly advise against using this model. It's really bad, complete nonsense.

I've also tried other models, but they were even worse.

So, I'm looking for a model that works well. I don't expect perfection, but at least translations that are complete and respect gender (masculine/feminine) properly.

I dont know korean, but I asked gpt about genders and it says its a gender neutral language just like english.
For example english just does not have enough info to always determine the gender correctly for example when translating to czech/slovak. (which has gender in many more words in a sentence)

Unless it was mentioned in some sentence that its talking about he or she, its just not possible to determine the gender. (and it might just guess it wrong, because it does not know who is speaking right now)
The language just does not provide enough info about genders.

The subtitles would always have to mention who is talking right now.

(I only use GPTs, so I cant recommend local ones)

[Neoony]

I was thinking

because default in instructions is "Keep any tags as they are."

That might be why you dont see the [태영] translated

it might think its a tag which should not be translated
try removing that from instructions

[Neoony]

The local models will most likely always be worse and have much more issues.

[Dydhzo]

Yes, I’ll test with Claude 3.5 Sonnet to see how it works. Thank you for your explanation, but I have a huge number of subtitles to translate, so the cost will add up quickly.
Yes, I’ll try modifying the default instruction values. The issue is likely coming from there—thank you.
I understand, and I’ll test with Claude/gpt4 to see how it performs. Thank you for your response.

[machinewrapped]

I haven't had much luck with local models, to be honest - it's a while since I've tried, so maybe the newer generation models are more reliable, but my early tests were discouraging.

I've started using Gemini 2.0 Flash Experimental as my goto translator recently - it's extremely fast, can handle large batches easily and seems to be smart enough to handle things like using provided names consistently that used to be a challenge for the smaller tier models. Best of all it currently seems to be free to use!

[Dydhzo]

The model I use the most is Qwen 2.5 32b. Overall, I find it very efficient. I just need to test without the 'Keep tags as they are' option as Neoony suggested - this will probably solve my issues.

I will also test Gemini 2.0 Flash Experimental which looks very interesting.

Thank you for your suggestions!

[Lod3]

Although I have not managed to install the regex module yet and get the CLI interface of Subtrans to use it
I managed to use the GUI and have had very good results with Doplhon-Mistral 2.6 more specifically
dolphin-2.6-mistral-7b-dpo-laser-GGUF
Compared to all of these

Dolphin-Mistral worked very good and kept working, it stopped without translated around 10 lines.
Subtrans seems to add a - in the translation field of the a line when a line is not translated.
And when restarting subtrans they get replaced with "none". I cleared those lines of content ( the - and none) and just clicked play again until they where translated.

sEZtt6S4rf.mp4

[Lod3]

franken-mistral-merlinite-maid-i1 also works good, faster compared to dolphin-2.6-mistral-7b-dpo-laser-GGUF
No weird symbols to clean out (dolphin mistral is made for coding primarily).
But the translation from Flemish to EN had more mistakes
Less lines are skipped I had to edit 1 lines manually and select 3 lines that where skipped to translate them again through the "Translate selection".
And 7 lines needed a translation fix
I used the same SRT file from Flemish to EN.

[Lod3]

asherah_multimodal_7b is VERY fast I think if it can be tweaked through the gui or in lm studio it can potentially be a good candidate.
It required some retranslating of 15 empty lines and 1 manual correction.
I dont know what the details of this model are if someone who knows more about this let me know I would love to know more.