This model is a fine-tuned version of google/gemma-3-4b-pt. As it is intended solely for text generation, we have extracted and utilized only the Gemma3ForCausalLM component from the original architecture.
Unlike our previous EEVE models, this model does not feature an expanded tokenizer.
Model Name:yanolja/YanoljaNEXT-Rosetta-4B-2510
Base Model:google/gemma-3-4b-pt
GGUF files
This folder contains ready-to-run GGUF files for llama.cpp.
BF16/YanoljaNEXT-Rosetta-4B-2510-bf16.gguf: full-precision reference model
Quantized variants (choose one based on your device and quality needs):
K-family: Q3_K_{S,L}, Q5_K_{S,M}, Q6_K, Q8_0
IQ-family: IQ2_{S,M}, IQ3_{XXS,XS,S}, IQ4_{XS}
For many types there are matching _IMX folders. Files there were produced with an activation matrix (imatrix.gguf) and usually offer better quality at the same size. In this release, IQ2_{S,M} and IQ3_{XXS,XS} are IMX-only.
Model Description
This model is a 4-billion parameter, decoder-only language model built on the Gemma3 architecture and fine-tuned by Yanolja NEXT. It is specifically designed to translate structured data (JSON format) while preserving the original data structure.
The model was trained on a multilingual dataset covering the following languages equally:
Arabic
Bulgarian
Chinese
Czech
Danish
Dutch
English
Finnish
French
German
Greek
Gujarati
Hebrew
Hindi
Hungarian
Indonesian
Italian
Japanese
Korean
Persian
Polish
Portuguese
Romanian
Russian
Slovak
Spanish
Swedish
Tagalog
Thai
Turkish
Ukrainian
Vietnamese
While optimized for these languages, it may also perform effectively on other languages supported by the base Gemma3 model.
How to use
Use a recent build of llama.cpp that supports Gemma 3 models. Pick any GGUF file from this folder (a quantized variant is recommended for most users).
bash
1# Example: use a Q5_K_M quantized file (adjust the path/model to your choice)2MODEL="path/to/YanoljaNEXT-Rosetta-4B-2510-q5_k_m.gguf"34# Build a formatted prompt using the included chat template roles5# (see release/YanoljaNEXT-Rosetta-4B-2510/chat_template.jinja)6read -r -d '' PROMPT <<'EOT'
7<start_of_turn>instruction
8Translate the user's text to Korean. Keep the JSON structure and keys.
9Context: Simple introduction about a tech company.
10Tone: Informative and helpful
11Glossary:
12- Yanolja NEXT -> 야놀자넥스트
13- travel industry -> 여행 산업
14Provide the final translation immediately without any other text.
15<end_of_turn>
16<start_of_turn>source
17{"company_name": "Yanolja NEXT", "description": "Yanolja NEXT is a company that provides cutting-edge technology for the global travel industry."}
18<end_of_turn>
19<start_of_turn>translation\n
20EOT2122# Run llama.cpp (adjust -n/-c/--temp as needed)23llama-cli -m "$MODEL" -p "$PROMPT" -n 64 -c 4096 --temp 0.7 -no-cnv
The model is optimized to output structured JSON for translations when appropriate.
Import any of the .gguf files into your GUI of choice (LM Studio, KoboldCPP, text-generation-webui) and select chat mode. The embedded template in the GGUF will be used automatically by recent tools.
Training Procedure
Training Data
The translation datasets were synthesized using fineweb corpora.
The model was fine-tuned on synthetic multilingual translation data to optimize performance across the supported language pairs.
Performance
Translation Quality Benchmarks
The following CHrF++ scores (WMT24++) demonstrate the model's competitive performance compared to other state-of-the-art translation models on English to Korean translation:
Model
CHrF++ Score (WMT24++)
google/gemini-2.5-flash-lite
35.23
yanolja/YanoljaNEXT-Rosetta-4B-2510
35.09
yanolja/YanoljaNEXT-Rosetta-12B
34.75
yanolja/YanoljaNEXT-Rosetta-20B
33.87
google/gemini-2.0-flash-001
33.81
openai/gpt-oss-120b
31.51
yanolja/YanoljaNEXT-Rosetta-4B
31.31
openai/gpt-4.1-nano
31.15
Qwen/Qwen3-235B-A22B-Instruct-2507-FP8
31.02
openai/gpt-oss-20b
30.56
google/gemma-3-27b-it
30.05
google/gemma-3-4b-pt
27.53
YanoljaNEXT-Rosetta-4B-2510 achieves competitive translation quality while maintaining the efficiency of a 4B parameter model.
Scores for the other language pairs can be found in the WMT24++ Evaluation Results.
Intended Uses & Limitations
This model is intended for translating structured data (JSON format) while preserving the original structure. It is particularly well-suited for tasks such as localizing product catalogs, translating hotel reviews, or handling any other structured content that requires accurate translation.
Limitations
The model is primarily optimized for processing JSON data.
Its performance on unstructured text or other data formats may vary.
In some cases, the model may produce invalid JSON, repetitive output, or inaccurate translations.
License
This model is released under the Gemma license, inherited from its base model, google/gemma-3-4b-pt. Please consult the official Gemma license terms for detailed usage guidelines.
Acknowledgments
This work was supported by the Korea Creative Content Agency (KOCCA) grant, funded by the Ministry of Culture, Sports and Tourism (MCST) in 2025 (Project Name: Cultivating Masters and Doctoral Experts to Lead Digital-Tech Tourism, Project Number: RS-2024-00442006, Contribution Rate: 100%).
Citation
If you use this model, please consider citing:
@misc{yanolja2025yanoljanextrosetta,
author = {Yanolja NEXT Co., Ltd.},
title = {YanoljaNEXT-Rosetta-4B-2510},
year = {2025},
publisher = {Hugging Face},
journal = {Hugging Face repository},
howpublished = {\\url{https://huggingface.co/yanolja/YanoljaNEXT-Rosetta-4B-2510}}
}
References
This work utilizes several models and datasets. We would like to acknowledge the original authors for their valuable contributions to the field.
@misc{gemma3,
author = {Google},
title = {Gemma 3},
year = {2024},
publisher = {Google DeepMind},
howpublished = {\\url{https://deepmind.google/models/gemma/gemma-3/}}
}
@misc{penedo2025fineweb2pipelinescale,
title = {FineWeb2: One Pipeline to Scale Them All -- Adapting Pre-Training Data Processing to Every Language},
author = {Guilherme Penedo and Hynek Kydlíček and Vinko Sabolčec and Bettina Messmer and Negar Foroutan and Amir Hossein Kargaran and Colin Raffel and Martin Jaggi and Leandro Von Werra and Thomas Wolf},
year = {2025},
eprint = {2506.20920},
archivePrefix = {arXiv},
primaryClass = {cs.CL},
url = {https://arxiv.org/abs/2506.20920},
}
@misc{lozhkov2024fineweb-edu,
author = {Lozhkov, Anton and Ben Allal, Loubna and von Werra, Leandro and Wolf, Thomas},
title = {FineWeb-Edu: the Finest Collection of Educational Content},
year = 2024,
url = {https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu},
doi = {10.57967/hf/2497},
publisher={Hugging Face}
}
