tmcleanerby terratra.com
…
◇ how it works

five steps. no surprises.

this page is the user-friendly walkthrough. if you want the engine internals, benchmarks, and full security posture — that lives at /under-the-hood.

━━ the 5-step flow

  1. 01
    sign in

    email magic link. no passwords stored, ever. lost access? request a new link.

  2. 02
    upload

    drag-drop a .tmx (up to 2 GB). uploaded directly to encrypted object storage via a one-time signed URL — bytes never touch our API node.

  3. 03
    preview · free

    we scan and show you exactly what will be removed: duplicates, junk, projected output size. nothing is cut until you proceed.

  4. 04
    pick a preset · then clean

    three named presets cover ~95% of jobs — Lenient / Balanced / Strict — one click, eleven options applied. fine-tune via the Custom rules panel if you need to. the engine streams through your file in two passes, never loading it into RAM.

  5. 05
    download

    one compressed bundle.zip (cleaned.tmx + removed.tmx + removed.csv). individual files also available. original upload is purged the moment processing finishes (unless you ask us to keep it 24 h so you can clean it again with other settings); downloads stay live for 1 hour, then the files are deleted.

━━ pick a preset · or fine-tune

three named cards above the rules panel — one click applies eleven options atomically. the right shape for ~95% of jobs. power users open the “custom rules” disclosure for the granular toggle panel underneath.

Lenient
keep more · QA & review

preserves untranslated entries, MT-flagged TUs, and short UI labels verbatim. Good for human review or pre-delivery inspection where you want the final say.

Balanced
default · LSP delivery

drops mechanical leakage (untranslated, empty target, MT-flagged) but doesn't run the strict QA validators. Sane defaults for delivering a TM to a client.

Strict
cut deep · MT training

everything Balanced does plus all QA validators on (placeholder mismatch, mojibake, length anomalies) and case-insensitive dedup for tighter merging. Right shape for MT-training datasets.

━━ what gets removed · or flagged

the engine ships eight cleanup classifiers, grouped into three categories. each is individually toggleable in the custom-rules panel; the presets above bundle sane defaults so you don’t have to pick.

duplicates
  • →exact matches by normalized source
  • →whitespace-only differences
  • →winner picked by Popular / Latest — or by your priority-author list
  • →manual per-group override always available in the preview
structural junk
  • →empty source / empty target
  • →single character · pure numbers
  • →URLs · emails · phone numbers
  • →pure tags · repeated symbols (--- / === / ***)
  • →suspect markup (closing-tag corruption from CAT round-trips)
translation-quality red flags
  • →untranslated (target = source · case + whitespace ignored)
  • →MT-flagged TUs (creationid / changeid contains MT engine signature)
  • →length outliers — target much longer OR shorter than source · NFC-normalised so CJK / Indic / Thai targets are treated fairly
  • →placeholder mismatch — {0} / %s / <ph/> count differs between source and target
  • →mojibake (encoding corruption — 'café' → 'café' style)
  • →brand / DNT terms missing in target — you list iPhone, HIPAA, etc. and we flag any TU where the term appears in source but not target

━━ priority authors

when duplicates conflict, you can rank the trusted translators whose variants should win. an ordered list (5–20 names typically) overrides the Popular / Latest strategy for any group where one of those authors has a variant; everywhere else falls back to your default. catches the senior-translator-vs-freelancer case that no major CAT tool solves at TM-cleanup time.

━━ what we keep, untouched

━━ privacy in plain english

full security posture (XXE handling, signed URLs, headers, retention policy) → /under-the-hood

clean a tmx now →