Prompt templates

A prompt template decides how each subtitle line is put to the translation model. Some models translate far better with their own prompt format: Tencent Hy-MT, for example, and Xiaomi MiLMMT, which doesn't translate at all with a general "you are a translator" prompt.

Pick one on the Model tab (Prompt Template), or leave it on Auto to choose by the model's name. The choice is saved with profiles and exported with them. Chinese subtitles are always converted to the chosen script (Traditional or Simplified), whatever the model writes.

The app's templates

These are the three files in the app's templates/ folder:

File	Template	Used automatically for	What it sends
generic.json	General	any other model	subtitle-translator instructions as a system prompt, with the earlier lines as conversation history
hy-mt.json	Hy-MT (Tencent)	models whose name contains hy-mt, hymt or hunyuan-mt	Tencent's official templates and sampling
milmmt.json	MiLMMT (Xiaomi)	models whose name contains milmmt	Translate this from Japanese to English: / Japanese: … / English:, greedy

You can edit them: the server uses the changed file from the next subtitle line on. If one of them is deleted or can't be read, the server falls back to its built-in copy of the original (and says so in its window), so translation keeps working. To undo your edits, delete the file and restore it from the release zip or with git checkout templates/<file>.

Adding your own

Put a .json file in the templates/ folder. The file name (without .json) is the template's id, so use letters, digits, -, . or _ (not auto, and not the three names above unless you mean to replace those). Files starting with _ are ignored, like _example.json: copy it to a new name to start. The server picks up new or changed files by itself; reopen the popup to see them in the menu.

{
  "name": "Formal English (Qwen)",
  "description": "Polite, formal wording",
  "match": ["qwen3.8"],
  "system": "Translate each subtitle line into formal, polite {target}. Output only the translation.{target_rules}",
  "user": "{text}",
  "history": true,
  "sampling": { "temperature": 0.2 }
}

Field	Required	Meaning
user	yes	The message with the subtitle line; must contain {text}
system	no	A system prompt, sent first
user_with_context	no	Used instead of user when there are earlier lines; must contain {text} and {context}
history	no	true sends the earlier lines and their translations as a conversation (user / assistant turns) before the new line
match	no	Words that make Auto pick this template when the model's name contains one (your templates are checked before the built-in ones)
sampling	no	Any of temperature, top_p, top_k, min_p, repeat_penalty, presence_penalty, frequency_penalty
name, description	no	Shown in the menu

Placeholders, filled in for each line:

Placeholder	Becomes
{text}	the subtitle line to translate
{target}	the subtitle language, e.g. Traditional Chinese
{source}	the speech language when known, e.g. Japanese, else the original language
{context}	the earlier lines, one per line as original => translation (empty when there are none)
{target_rules}	extra rules for some subtitle languages (Traditional Chinese: Taiwan usage only), else empty

Any other text in braces is sent as it is. A file that can't be read, or that breaks one of these rules, is skipped with a message in the server window saying why.