Local model experiment · October 9, 2026

What My Local Translation Candidate Tests Actually Tell Me

I'm learning how to train smaller models and build the software around them. Local captions and translation give that learning a purpose: useful software people can actually run.

The wider goal is a video-editing suite with local captions and translations. It is still development work, not a claim that a new app release is ready. Most of my experiments have involved Qwen, but the public candidate discussed here comes from Tencent Hy-MT2-1.8B. It is a separate artifact from my Qwen 1.5B work.

Start with the exact candidate

The model card describes targeted preference adaptation at step 64, followed by merging and conversion through the official LiteRT export path. It identifies a Q8 model.litertlm artifact and its SHA-256. The documented native interface returns one translated sentence; it is not a Qwen JSON caption-batch interface. Model family, prompt format and runtime all matter when building the surrounding software.

What was checked

The card reports 432 native Linux cases passing its recorded JSON/fence/HTML instruction-execution gate. It also reports 76 selected frozen cases on each of two Android phones, Seeker and Samsung, with no mechanical JSON/fence/HTML failures, plus cancellation and fresh-process recovery checks. The phone cases use English source across 18 target languages: 72 command-format cases and four ordinary controls.

These are publisher-reported checks of a specific artifact and setup. The prompt policy treats spoken source and neighboring context as data, not instructions. Passing a mechanical gate means the recorded output avoided the checked failure patterns; it does not establish that every translated sentence is correct.

What is still open

Meaning was not independently graded.
The recorded phone coverage does not establish translation quality in every language direction.
Model-only tests do not establish production app integration or final APK/AAB acceptance.
The card notes that plain “ok” may remain unchanged.

No private user-caption data was used in the recorded tests, according to the card. Read its current evidence and limitations alongside any result; a public candidate is not a finished editing suite.

Why this helps me learn

My questions are about training instructions, response formats and how a model and its harness can be developed together. A useful next evaluation would inspect meaning separately from mechanical output: negation, speaker roles and uncertainty, reviewed by people who understand the language pair. That is a proposed next check, not a result already established by this candidate.

Human First AI means keeping the speaker's meaning and the user's control central to that work. I want critical feedback on the documented limits and reproducible failures. More experiments are continuing behind the scenes; I will distinguish what has been tested from what I hope to build.

This note was prepared with my AI agent from my stated project goals and the public model card. It summarizes reported evidence; it is not an independent replication.