Repository navigation
How It Works
Here's what happens between opening a folder and getting translated files, in order. You tick the episodes and press one button; the rest is automatic.
When you open a folder, Polygluttony reads the .ass files in it, without using any AI:
- Source language: it counts which writing system the dialogue uses (Han characters, Hangul, kana, Cyrillic…) and picks the most common one.
- World type (Chinese only): it counts genre keywords such as 修仙 and 金丹 (xianxia), 江湖 and 内力 (wuxia), or 皇帝 and 朝廷 (historical). No hits means modern. The world type decides which terms the glossary focuses on and sets a matching tone.
You can change both with Edit next to the settings line on the Episodes page. Episodes without a translation are ticked, ready for the next run.
A glossary is a list of names and terms with one agreed translation each, so 叶长生 is "Ye Changsheng" in every episode.
While Build a glossary first is ticked (it is by default), the gold button builds the glossary before translating, or adds the new episodes to the one you have. It uses up to three AI steps:
- Find names reads every line of the ticked episodes and pulls out characters, places, techniques, items, organizations and cultivation terms.
- Tidy up fixes inconsistent spellings and drops terms too generic to be useful. It's on for this step.
- Look up names online finds the names fans already use for the show, so your glossary matches the wikis. It's optional and only runs when you start it from the Glossary page.
Adding episodes only adds new terms. Terms already in the glossary, including the ones you edited, stay as they are.
The result is a glossary.json in your folder, plus a hidden .polygluttony folder that remembers where each term came from, which episodes the glossary has read, and its last 10 versions. More: The glossary.
Each episode is cut into batches of lines: 100 per request by default, 40 for models on your own computer. Each batch goes to your Translate model together with:
- the glossary terms that appear in those lines, marked as required translations;
- the last few lines before the batch, so the model has context;
- your tone.
Before sending, Polygluttony takes out the formatting tags ({\an8}, {\pos(…)} and so on) and tags every line with a numbered marker. When the reply comes back it checks the markers. If the model dropped, merged or reordered lines, the good part is kept and the rest is sent again, in smaller batches if needed. The tags then go back where they were.
Afterwards, any line still containing untranslated source text gets a quick clean-up pass.
If a provider asks Polygluttony to slow down, it waits and says so on the episode's row, then carries on.
When an episode is translated, Polygluttony checks it in two stages:
- An automatic comparison, with no AI. It compares source and translation line by line: questions that stopped being questions, glossary terms that went missing, lines that are far too short or long.
- An AI spot-check. Your Verify model reads a sample of lines from the start, middle and end of the episode and flags any translation that doesn't say what the original says.
Flagged parts are translated again, up to twice. Then every line is checked against the glossary one last time. A line where a name doesn't match is listed for you, but isn't translated again.
Anything still wrong is listed as issues, and the file is saved with .warning. in its name. There's no quality score. More: Reading the results.
A finished episode is written next to its source, as Episode 01.eng.ass or Episode 01.warning.eng.ass. Writing one removes the other. An episode is only written once it's finished, so a stopped run never leaves half a file behind.
Polygluttony also keeps each episode's result in its own data folder: the lines, the open issues, and how the run ended. That's what lets you review an episode later, even after a restart, and what lets a stopped run carry on where it left off instead of starting over. See Privacy & your data for where it lives.
Each of the five jobs above is a "step" on the Models page. They can all use one model, or each its own.
| Step | Part of | What it does |
|---|---|---|
| Find names | Glossary | Finds names and terms in your subtitles |
| Tidy up | Glossary | Fixes spellings and drops generic terms |
| Look up names online | Glossary | Finds the names fans already use (optional) |
| Translate | Translation | Every line, plus clean-up passes |
| Verify | Translation | Checks each episode and redoes lines that slipped |
Only dialogue text, glossary terms, your tone and the prompt, plus the show's name if you use Look up names online. File names, folder paths, styles, timings and fonts never leave your computer. Details: Privacy & your data.