Repository navigation
v1.14.3
What's new since v1.14.2
Back up your database before upgrading if you translate with Ollama or an OpenAI-compatible backend — this release deletes wrong-language translation-memory entries. See Upgrade notes below.
Fixed
- Translations go in the direction that was asked for. Every Ollama and OpenAI-compatible request told the model to translate "from English to German", whatever the job wanted. A German subtitle sent for an English target came back as reworded German and was saved as
.en.srt. Prompts now name the requested languages; German targets keep the exact prompt the fine-tunes were trained on. - A machine translation is never used as a translation source, and a genuine subtitle that later replaced one counts as genuine again.
- Replacing a machine translation never leaves an episode worse off. Approving a pending original could trash the machine translation and then fail to install the original, leaving no subtitle; a translation in another format than the original was not removed. The translation is now set aside and restored unless an original actually landed on disk.
- Duplicate detection belongs to the episode. A subtitle byte-identical to one from another episode in the same season folder was skipped as a duplicate, and the episode stayed without a subtitle.
- Embedded subtitle tracks count in standalone mode (#36) and show in the episode list (#35) instead of "No subtitle found". Forced and "Signs & Songs" tracks no longer count as full subtitles. In standalone mode an SRT satisfies the language when SRT→ASS upgrades are off.
- No more stalls behind open browser tabs. Socket.IO polls held the server's four request threads; a few open tabs could make even
/healthwait 24 s. The server runs 100 threads now (one process): 24 s → 42 ms in the same situation. - Radarr movies: subtitle list, upload and combine work (previously 404), and the movie page shows the real missing count and video file.
- Data cleanups reach databases without migration history (typical for SQLite installs created by Sublarr itself), once each.
- Page size outside 10–200 is rejected; phone layout of Library and Wanted no longer overflows sideways; the optional RQ worker has a real healthcheck.
Added
- Approve pending originals in bulk. Select originals (or all) and approve them together; they install one after another in the background with progress in the dialog, which is wider and no longer scrolls sideways.
scripts.repair_wrong_direction_mt— lists (dry run) and trashes machine-translated subtitle files written in the wrong language, then searches those episodes again.
Changed
- Removed an unused second Ollama translation client (#29).
Upgrade notes
- Migration
tm4_wrong_directiondeletes translation-memory entries from the Ollama and OpenAI-compatible backends whose target language differs from the configured one. On databases without migration history,tm1_strip_breaks,tm2_drop_same_langandtm3_strip_softalso run once. - Repair files already written in the wrong language (dry run first):
docker exec -it sublarr gosu sublarr python -m scripts.repair_wrong_direction_mt docker exec -it sublarr gosu sublarr python -m scripts.repair_wrong_direction_mt --apply
- If your compose sets its own
command:, use--threads 100there too. If you run the RQ worker from your own compose, add the healthcheck fromdocker-compose.redis.yml.
Tested via
1.14.3-rc.1 … rc.4 on a production-data mirror, an untracked SQLite beta instance and an independent VM test (rc.2, rc.3, rc.4). On the reference install: 70 542 poisoned memory entries removed, 342 wrong-language subtitle files repaired.
Deployed: ghcr.io/abrechen2/sublarr:1.14.3 (amd64 + arm64; also :latest, :stable, Docker Hub)
Docs: https://sublarr.de/docs/