cstr/opus-mt-de-ar-GGUF

Opus-MT German → Arabic — GGUF (ggml)

GGUF / ggml conversion of Helsinki-NLP/opus-mt-de-ar for use with CrispStrobe/CrispASR (--backend marian).

A small (~80M parameters, 6+6 layers, d=512) MarianMT model for one language pair. In CrispASR these are the fastest translators for live transcription + translation (--live-translate).

Licence and attribution

CC-BY-4.0. The weights are the opus-2020-07-03 release of the OPUS-MT project (Jörg Tiedemann and Santhosh Thottingal, University of Helsinki; OPUS-MT — Building open translation services for the World, EAMT 2020), trained on OPUS data. The project distributes its pre-trained models under CC-BY 4.0 (Opus-MT README, OPUS-MT-train README); that statement, not the tag on an individual model card, is what this repository follows. This is a format conversion: the weights are unchanged at f16 and quantized at q8_0. Redistribution requires attribution to the OPUS-MT project.

Files

File	Size	Output vs. the reference implementation
opus-mt-de-ar-f16.gguf	159 MB	8/8 sentences identical greedy, 8/8 with beam 4 — recommended for this pair; -m opus-mt-de-ar downloads this file.
opus-mt-de-ar-q8_0.gguf	87 MB	3/8 sentences identical greedy — the others differ from the reference; not recommended for this pair.

"Reference" is MarianMTModel.generate from Hugging Face transformers on the original checkpoint, on 8 test sentences; input token ids were identical for 8/8.

Quick start

git clone https://github.com/CrispStrobe/CrispASR && cd CrispASR
cmake -B build -DCMAKE_BUILD_TYPE=Release && cmake --build build -j

# Text → text. -bs 1 is greedy; without it the checkpoint's own beam size is used.
./build/bin/crispasr --backend marian -m opus-mt-de-ar -sl de -tl ar -bs 1 \
    --text "Guten Morgen und herzlich willkommen zur heutigen Sitzung."
# صباح الخير ومرحبا بكم في جلسة اليوم

# Live: microphone in, transcript + translation out, sentence by sentence
./build/bin/crispasr --live-translate -l de --tr-tl ar \
    -m auto --backend parakeet --translate-backend marian

-m opus-mt-de-ar downloads opus-mt-de-ar-f16.gguf on first use (unlike the other pairs, whose q8_0 matched the reference on 6-8 of 8 sentences). The recogniser in the live example must support German.

Notes

• One model per direction; the opposite direction, where one exists, is cstr/opus-mt-ar-de-GGUF.
• Live mode always decodes greedy.
• Literal </s>, <unk>, <pad> in the input are treated as ordinary text here (the reference treats them as special tokens).

Conversion

python models/convert-marian-to-gguf.py --input <dir>/opus-mt-de-ar --output opus-mt-de-ar-f16.gguf
./build/bin/crispasr-quantize opus-mt-de-ar-f16.gguf opus-mt-de-ar-q8_0.gguf q8_0
python tools/marian_parity.py --hf-dir <dir>/opus-mt-de-ar --gguf opus-mt-de-ar-f16.gguf --lang en --tgt ar --sentences <file>