Try TM Cleaner on a sample translation memory

The demo runs the real preview on a sample file: an English → French translation memory of about 270 segments from a made-up notes app, with the problems real TMs collect. Nothing on the next screen is staged. It's the same preview, the same engine and the same controls you get with your own file.

━━ what's in the sample

• ◇Duplicates: the same English sentence stored up to 12 times, sometimes with different translations.
• ◇One of each of the 14 junk kinds: a web address, an email, a mailto: link, a phone number, an amount, a hash, a file path, a version number, a placeholder-only pair, a tag-only pair, broken markup, punctuation, a separator line and symbol noise.
• ◇Untranslated and empty translations, and pairs a machine produced, recognised by their author stamp (DeepL-API, MT).
• ◇Three translations with broken characters, like “stockées” for “stockées”.
• ◇Placeholders dropped from the translation, translations far longer or shorter than their source, and the product name translated by mistake.
• ◇Five translators and dates from 2016 to 2025 to filter by, including an old TM import.

━━ things to try on the preview

1. 01 · Switch presets

   Balanced is selected. Click Strict and the counts change at once: it adds the quality checks (length, placeholders, broken characters) and treats “Sign in” and “SIGN IN” as duplicates. Lenient keeps untranslated and machine-translated pairs.

2. 02 · Choose which junk goes

   Under Junk & broken pairs → Choose which kinds, switch a kind off and its segment is kept. The sample has one of each of the 14 kinds.

3. 03 · Repair instead of remove

   Under Quality checks, turn on Repair broken characters. “stockées” becomes “stockées”, and the three repaired pairs are listed before → after.

4. 04 · Decide a duplicate

   In step 03, Review before you clean, open “Click Save to apply your changes.” Its five versions differ in letter case, spacing, punctuation or wording, and the difference is highlighted. Keep another version, or remove the whole group. Under Duplicates, switch from Most common to Newest and the group keeps j.martin's 2025 wording.

5. 05 · Merge letter-case variants

   Turn on Ignore letter case: “Sign in”, “Sign In” and “SIGN IN” become one group.

6. 06 · Trust a translator

   Add marie.dubois under Preferred translators. Her version of “An error occurred. Please try again later.” now wins, although another one is more common.

7. 07 · Protect a name

   Add Kestrel under Protected terms. The pair that translated the product name as “Crécerelle Notes” is removed.

8. 08 · Filter by translator or date

   Under Translators & dates, exclude old-tm-import, or everything before 2020.

9. 09 · See everything that goes

   Step 03 lists what's being removed, by reason. Keep one row, every row shown, or a whole kind at once. Browse every removed segment lists every decision, not just a sample.

10. 10 · Clean and download

    No account needed for files up to 50 MB (a free sign-in by email link takes it to 2 GB). You get the cleaned TMX, the removed segments as a TMX, and a CSV of every removal and its reason. Tick the options under Download files for Excel and TSV copies too.

━━ what happens to the sample file

It's uploaded and handled like any other file. If you don't clean it, it's deleted when you press Cancel, or automatically within about a day. If you clean it, it's deleted as soon as cleaning finishes (unless you ticked Keep my original for 24 hours), and the download links work for one hour. Privacy policy.