Skip to content

Releases

v0.1.0 - Initial model release

Choose a tag to compare

@slayyyyer slayyyyer released this 14 Nov 17:06

Initial public pilot release · Windows 10/11 x64 · English ↔ French

Transcend Translator is an offline desktop translator built around a small
neural machine-translation model trained from scratch for this project. This is
a research pilot: it is useful for getting the gist of a short sentence, not for
production, legal, medical, safety, financial, or customer-facing work.

Download

File Size Purpose
Transcend-Translator-Setup.msi 262,136,208 bytes (about 250 MiB) Windows installer; includes the model
Transcend-Translator-Setup.msi.sha256 97 bytes SHA-256 checksum file
Transcend-Model.zip 145,549,835 bytes Exported model package for inspection; not needed to install

The model is bundled in the installer. You do not need Python, a separate model
download, an account, or a network connection after installation.

Install

Requirements: Windows 10 or 11, 64-bit x64. The installed application uses
about 360 MB; allow about 700 MB free during installation. ARM64 is not
validated.

  1. Download the MSI and, if you want to check it locally, the .sha256 file.
  2. Verify the MSI's SHA-256 against the value below. Do this before running it.
  3. Run Transcend-Translator-Setup.msi.
  4. If Windows shows “Windows protected your PC” or “Unknown publisher”,
    choose More info → Run anyway. The installer is not code-signed;
    this warning is expected for an unsigned pilot, not a malware detection.
  5. No administrator is required. Setup is per-user, installs under
    %LOCALAPPDATA%\Transcend Translator, and adds a Start-menu shortcut.
  6. Launch Transcend Translator. The model loads locally and the app works
    offline.

Smart App Control on a fresh Windows 11 installation may block unsigned
installers entirely. A code-signed build is the permanent fix; do not disable a
Windows security feature just to run this pilot.

Verify the download

Open PowerShell in the folder containing the MSI and run:

Get-FileHash .\Transcend-Translator-Setup.msi -Algorithm SHA256

The SHA-256 must be:

9666e1991e331c81bb7b0eb757dd4adfc6f9e77aaef581d7677f607860973620

This is the same hash stored in the attached
Transcend-Translator-Setup.msi.sha256 file. If the hashes differ, delete the
download and do not run it.

If you also downloaded the optional model archive:

Get-FileHash .\Transcend-Model.zip -Algorithm SHA256

Its SHA-256 must be:

c4eb95f4b86f47ad4b6ea82923c3b7c404a886f5c796938cecec5577198d3c76

Build attestation

v0.1.0 predates the repository's GitHub artifact attestations. There is no
attestation for this MSI; the SHA-256 above is its published provenance. Builds
from v0.1.1 onward are attested by the Windows installer workflow. With the
GitHub CLI installed, verify a future MSI with:

gh attestation verify ./Transcend-Translator-Setup.msi --repo slayyyyer/transcend

An attestation binds the MSI to the repository, workflow, commit, and tag. It
does not replace code signing.

Uninstall

Open Settings → Apps → Installed apps → Transcend Translator, then choose
Uninstall.

Model provenance

Item Value
Architecture 19,825,920 parameters; 6 encoder + 6 decoder layers; d_model 320; 5 attention heads
Tokenizer 8,000-token byte-level BPE
Training From scratch; 350M-token stage resumed to an 800M-token budget
Shipped checkpoint Dev-selected step 77,000 of 91,691; 671,795,250 tokens seen
Model ID scaled_full@step77000#a3110fb2
Training data Europarl v7 + WMT14 News Commentary v9; 2,101,321 training pairs
Not used by this model Tatoeba and ParaCrawl
Runtime limits 128 source tokens; 136 generated target tokens
Encoder SHA-256 a3110fb27f91fc1e8ac616d3b29c072c6361b0153ebcc7d14d07f43e64614539
Model archive SHA-256 c4eb95f4b86f47ad4b6ea82923c3b7c404a886f5c796938cecec5577198d3c76

The model was trained from random initialization on English–French Europarl and
WMT News Commentary data. It does not use pretrained weights, Tatoeba, or
ParaCrawl. The attached archive contains the exported model package; the
installer already bundles the model, so downloading the archive is optional.

How well it works

These are automatic corpus scores for the shipped step-77,000 checkpoint on
newstest2014, using greedy decoding:

Direction BLEU chrF++
English → French 25.02 51.07
French → English 23.33 50.60

BLEU and chrF++ are corpus-level automatic metrics, not percentages of correct
sentences, and they do not establish that an individual translation is usable.
No qualified human evaluation of this checkpoint has been performed.

Known limitations

  • Pilot quality, not production quality. Do not use the output for legal,
    medical, safety, financial, or customer-facing content.
  • Outputs can be fluent-looking but wrong. The model can omit content,
    choose the wrong word, invent morphology, or repeat phrases.
  • One short segment at a time. Inputs above 128 source tokens are rejected
    with an error rather than truncated. Generated output is capped at 136 tokens.
  • English and French only. There is no automatic language detection,
    document context, HTML handling, or speech input.
  • Domain mismatch is likely. Training data is parliamentary and news text;
    conversational language, informal register, and modern named entities are
    underrepresented.
  • Automatic metrics only. The shipped checkpoint has not received a
    qualified bilingual human evaluation, and no COMET or full human review was
    performed.
  • No quantified safety audit. Bias, toxicity, and privacy were not measured.
    The training data includes political content and personal names inherited
    from the source corpora.
  • The installer is unsigned. Windows SmartScreen may warn, and Smart App
    Control may block the installer entirely.
  • Model and data terms are research-only. The model artifact inherits the
    Europarl v7 and WMT14 News Commentary v9 corpus constraints. No commercial or
    redistribution rights are asserted for the model or corpora. The source code
    is licensed separately; see NOTICE, LICENSE, and
    docs/DATA_CARD.md.

Release provenance note

v0.1.0 predates embedded installer build metadata and artifact attestations.
The published MSI does not contain app/build-info.json; the checksums and
model record above are the published provenance for this release.

More information