Do you have a dumb question that you're kind of embarrassed to ask in the main thread? Is there something you're just not sure about?
This is your opportunity to ask questions. No question too simple or too silly.
Culture war topics are accepted, and proposals for a better intro post are appreciated.
Jump in the discussion.
No email address required.
Notes -
What is the best LLM for translating to and from CJK? Please give an answer for both online stuff (Gemini, Claude, Copilot) and local stuff.
I tried Mistral 7B, but it seemed to be doing a much worse job than Gemini.
I would assume that most SOTA LLMs would be nigh interchangeable when it comes to translation tasks. When it comes to local models, well, what can your pc handle? The new Qwen is crazy sparse, but it's still beyond most consumer hardware.
I have 16 GB of RAM, a GTX 1660 Super, and a Ryzen 5 3600 CPU. Far from ideal hardware, I know. Mixtral 7B runs fine, though.
My experience with 7B models is that while it's amazing that they can do some useful tasks, they're severely limited. I mean, in some ways, they're better than GPT 3 or 3.5, but still far from reliable. Your specs aren't good enough to run larger models, I'm afraid, so you'd be better off using the bigger online models/closed-source models. At the bare minimum, you'd need a GPU with more VRAM (some manage barely usable results on CPU, but it's a PITA). I believe that some Gemma models might fit on your system, and probably do better than Mistral 7B, but at the end of the day, you need better.
If money is an impediment, Google is practically throwing away FLOPs on AI Studio. Sign up, and you can use Gemini 2.5 Pro without running into rate limits. I'm very confident that translation tasks are piss easy for it, and I've used it to translate hundreds of chapters of Chinese novels with satisfactory results.
Wow, you too? I find Gemini translates stuff really, really well. Or at least, I think it does, I cross-reference everything I put into it with Google Translate and sometimes it will apparently misunderstand something, or I'll open up a new private tab and have it translate it back to me and it's come out a little differently than I wanted it. I was told DeepL was supposed to be the best, but I somehow doubt it could be better than Gemini. I could be wrong.
The power of AI might shake up the closed nature of different language spheres of the Internet someday, I think. I had the fanciful idea that humans in other cultures would see things significantly differently from me, but I was surprised by how similar the lines of thought run. In almost every non-Arab country in the world, it seems that traditionalism is fighting a bloody battle against liberalism, and mostly losing.
Most recent models do really well! Translation is effectively solved since GPT 4, which was ages ago.
DeepL is good, but they're still better.
If you've successfully translated Chinese novels, can you recommend a model for translating from Chinese, but for someone that lives in China and doesn't want to use a VPN? I think Gemini is off-limits for them, and ChatGPT fucks with the writing style too much.
I believe that you should have access to Kimi K2, GLM and Qwen, which are the current best Chinese models. Can't imagine they'd be banned in China.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link