|
|
|
|
|
|
by Alex-Programs
394 days ago
|
|
|
You might find my LLM translation research useful:
https://nuenki.app/blog Your best bet is probably: - Produce the sentence with Qwen 3. I didn't test down to 8B, but its 32B variant does reasonably well (see https://nuenki.app/blog/claude_4_is_good_at_translation_but_...) and Chinese models are better at Chinese in general - Then prompt Qwen 3 again, this time telling it to critique the translation and improve it LLMs tend to be better at post-critique than generation, though I can't say I've tested with models that small. You may find https://nuenki.app/translator interesting. You might also be able to use some hideous distil of Deepseek V3? |
|
|