luweigen/eurollm-1.7b-instruct-android

EuroLLM-1.7B-Instruct — GGUF for Android

Source

• Original model: utter-project/EuroLLM-1.7B-Instruct (Apache-2.0), 35 languages.
• Converted with llama.cpp b11700 (convert_hf_to_gguf.py to f16, then llama-quantize).
• This repository is used by the native Android app of the vibevoice-android project for sentence-by-sentence translation of speech recognition output. The app downloads the files and checks each one against a sha256 value compiled into the APK. See the original model card for intended use, limitations and the license terms.

Files

File	What
eurollm-q8_0.gguf	Q8_0 (1.76 GB). On FLEURS (100 sentences, greedy) it matches f16 within noise
eurollm-q4_0.gguf	Q4_0 (1.0 GB). Noticeably weaker when Finnish or Chinese is the target (en→fi −5, fi→zh −8 BLEU vs f16)

No other changes; the tokenizer and chat template are the original ones.

Where the files apply

• llama.cpp (CPU, OpenCL, Hexagon backends). The app uses Q8_0 on phones with ≥ 12 GB of RAM and Q4_0 otherwise.
• Use greedy decoding with the model card's ChatML prompt; sampling halves the BLEU score in our tests.

Usage and code

To be added.