GoePT-1-20M
GoePT-1-20M is a 19,943,808-parameter Marian encoder-decoder that translates Standard German into Alman, a constructed dialect of German that removes grammatical gender and most case inflection.
This repository is the browser deployment package: int8-quantized ONNX weights plus everything needed to run the model client-side in single-threaded WebAssembly. No server required.
Contents
| Path | Description |
|---|---|
model/onnx/encoder_model_quantized.onnx | int8 encoder (17.5 MB) |
model/onnx/decoder_model_merged_quantized.onnx | int8 merged decoder with past-KV caching (15.4 MB) |
model/tokenizer.json | SentencePiece unigram 16k tokenizer with byte fallback |
model/config.json, model/generation_config.json | Marian config, greedy decoding |
translator.mjs | self-contained ESM translator module (bundled Transformers.js + ONNX Runtime Web) |
wasm/ | single-threaded ONNX Runtime WASM binary and glue |
browser.json | Per-file SHA-256 manifest, release checks, and current benchmark scores |
Architecture
6 encoder layers, 1 decoder layer, width 384, FFN 1536, 8 attention heads, 1024 source and target positions, tied embeddings, greedy decoding. Distilled from a ByT5-Base teacher on a 9,999,555-row German→Alman corpus.
Quality
- AlmanBench acceptance: 894/1029 (86.88%) in-browser
- Held-out eval exact match: 2540/3204 (79.28%) in-browser, 2538/3204 (79.21%) native
- Native↔browser output agreement: 99.71% (AlmanBench), 99.15% (validation)
- 2,018-word page translated in 6,428 ms, single-threaded WASM (Chromium 150); ~19.8 sentences/s warm, ~1.0 s cold init
All release gates in browser.json passed.
See the AlmanBench leaderboard, per-case results, and model announcement. You can use the model in the translator or Almanpedia.
Usage
const { createTranslator } = await import(
"https://huggingface.co/osolmaz/GoePT-1-20M/resolve/main/translator.mjs"
);
const translator = await createTranslator();
const [alman] = await translator.translate(["Der Hund ist im Garten."]);
The module resolves model/ and wasm/ relative to its own URL, so it also
works self-hosted: mirror the repository contents to any static origin.
Provenance
Mirrored from the frozen release candidate
osolmaz/alman-student-spm16k-base-10m-h50-onnx@8049cbd9cee1f9f417540e14df788022e9e7a5b7
(program ALMAN-TS-20260723-01); file-level SHA-256 digests in browser.json.
Native (PyTorch) weights: osolmaz/alman-student-spm16k-base-10m-h50.