Request Tool What local LLM model do you use in translation/translator like LunaTranslator?

CYBERBRID

New Member
May 10, 2024
4
2
I found this a few weeks ago, it can use hook/OCR(local or online)/copy text to translate game text (or other paragraphs if you like)

It support online api (google/deepL...),
online LLM api (ChatGPT. Gemini, Clade, Cohere)
offline translation (Sugoi, and some developted Chinese engine?)


One of the best thing is, it also supports local LLM, as long as you use Openai endpoint -which is quite easy if you use LM Studio to run local server.
屏幕截图 2024-12-17 203641.png
Just an example, the bar can move anywhere, and you can set other languages

So far some LLM models I used are:
1. -
7B -based on Qwen1.5/2.0/2.5(multiple version), especially for Japanese-Chinese translation, maybe also can be used for English translation?
It also has bigger version like 14B/32B if you have a powerful GPU (16GB Vram or more)

The Quantization method is Q6_k top for 7B and 13B, (2B or 1.5B is FP16, 32B is Q4_km)
For me 7B has already can complete with deepl, especally the model can read the context and it is totally local.
You can also use the GPT dictionary (not sure if it supports other llm model well)


---------
2.

This model is famous for translation, quite famous on reddit, and it supports 23 languages.
There are 8B/35B version. There is censorship but not strong.

Although there are new model , but the censorship is much more stronger than the previous version.


-----------
3. Llama 3.1

Not good enough compared to aya, I haven't try 70B (not to mention 405B), or Llama3.3 70B, my computer cannot run such a big model. And there's also strong censorship.

4. Qwen2.5
I didn't use it to translate sentence by sentence. However when I try to use it to translate paragraphs/article, it always tries to summarise the meaning and make it short. Using prompt like don't summarize can make it better, but the skipping still exised.

Some models I haven't tried or used for long time but recommend by others
5. Mistral
-8B
-Nemo (12B, nv has 8B version)
-Small (22B)
6. Gemma-2
7. Hermes-3


So any models or their finetune version that amazed you? Or the models I didn't mention?
 
Last edited: