Which LLMs are good for text translations?

Posted by badabimbadabum2@reddit | LocalLLaMA | View on Reddit | 12 comments

Hi, Which LLMs can translate text to popular languages? Does any locally run llm come close to chatGPT4 levels ? How much GPU vram do I need to have a good translator for production use?

Reply to Post

12 Comments

[-]

alexeir@reddit

I open-sourced 12 language models (24 language pairs) with a good quality on-pair with ChatGPT4. No any hallucinations. They work very fast even on mobile and have 120 mb and only 350 mln parameters. You can download them here: [https://github.com/lingvanex-mt/models](https://github.com/lingvanex-mt/models) You can test quality here: [https://lingvanex.com/translate/](https://lingvanex.com/translate/) if you need other languages Write me to [alexeir@lingvanex.com](mailto:alexeir@lingvanex.com)

[-]

Eugr@reddit

Depends on the language and topic. I've had good results with Mistral Nemo and Qwen2.5:32B. Gemma2 is good, but very censored.

[-]

Chaosdrifer@reddit

A specially trained/fined-tuned model for a certain language mostly would perform better than just the regular model. Like qwen2.5 would beat llama 3.1 at the same parameter size for translating between English and Chinese . But a model that’s tuned for Korea would do better for Korea translation. So if you want the best translation, you would want a fine tuned model that’s specific to the language you need and supplemented by a good glossary pertaining to the domain of the translated materials. And on top of that , a multi-agent approach to have the LLM to post editing , error check, proof-read, and finally human review/feedback.

[-]

Terminator857@reddit

You might want to test. Sometimes I find that gemini does better than gpt4. Some models do better in certain languages. For example mistral does very well in French. 7 months ago llama did o.k. with arabic.

[-]

Down_The_Rabbithole@reddit

Funny how mistral does well in french

[-]

OcelotOk8071@reddit

Mistral is a French company

[-]

mylittlethrowaway300@reddit

I'll ask a newbie question here without making my own post. Are transformer models that are encoder/decoder (like T5 or BART) better at translation than decoder only models (like GPTs, llama, Mistral)?

[-]

FullOf_Bad_Ideas@reddit

I think decoder only models received so much more attention and were trained on so much more data that they are much better at translation now. It should be possible to make better encoder/decoder models but money isn't there now.

[-]

mylittlethrowaway300@reddit

Gotcha. That makes a ton of sense. How much processing power went into training something like GPT2? And I'm guessing there's an optimal amount of training for a certain parameter size (and architecture)?

[-]

FullOf_Bad_Ideas@reddit

Here's a chart I found https://res.cloudinary.com/lesswrong-2-0/image/upload/f_auto,q_auto/v1/mirroredImages/iQx2eeHKLwgBYdWPZ/obwvtsaozxvglzbhuvgy There's compute optimal training and inference optimal training, generally most models nowadays are closer to inference optimal than compute optimal.

[-]

ReddaHawk@reddit

llama-3.1-nemotron-70b made by nvidia also is pretty good

[-]

Premium_Shitposter@reddit

Aya and Gemma2 are pretty good for translation