Semi-automatic translation (impersonation)

I am looking for what could be called semi-automatic, guided, or assisted translation. I want to manually modify some part of the translation and have the rest of the text translated accordingly. This would be used to encourage models to use my preferred vocabulary and style.

In general-purpose LLMs, this is called impersonation. For example, in Open WebUI, you can edit LLM’s response, remove everything that follows the edit, and click “continue” button to have the LLM regenerate the rest of the response accordingly. LLMs are good at in-context learning, so this technique is usually effective. Unfortunately, no local model I have tested is good enough at my language pair. Cloud models are clearly better at translation, but they do not support impersonation. LLMs can be also instructed, but IME even cloud LLMs thoroughly ignore instructions when translating.

Is there a way to impersonate/guide translations in LibreTranslate or the underlying Argos Translate? Are the underlying models even capable of maintaining translation consistency or would they ignore all guidance? If this is impossible or out of scope for LibreTranslate, are there opensource alternatives? Is there a different way to translate semi-automatically?

LibreTranslate has no mechanism to do this.

I wouldn’t rule it out of scope, it could be a cool feature to have/add.

OpenNMT and CTranslate2 have some support for customizing and constraining decoding. We currently don’t really expose this functionality in Argos Translate and LibreTranslate though.

Here are some links you can look into: