Monarda G4 E4B GGUF
Monarda is an offline Korean-English translation post-editor for Liquid Translate. It reads the original text and a Bergamot draft, then returns only the corrected translation.
Files
monarda-g4-e4b-base-q8_0.gguf: Q8_0 conversion of google/gemma-4-E4B-it for llama.cpp.monarda-translation-correction-focus-f16.gguf: Monarda translation correction LoRA adapter.
Both files are required. Load the base with the adapter:
1llama-server -m monarda-g4-e4b-base-q8_0.gguf \
2 --lora monarda-translation-correction-focus-f16.gguf \
3 --ctx-size 8192
Intended use
- English to Korean and Korean to English translation correction.
- Offline use after the files have been downloaded.
- Post-editing a machine-translation draft while preserving names, numbers, formatting, and technical terms.
The model is not intended to provide factual, legal, medical, or safety-critical advice. Users should review important translations.
Training and data
The adapter was trained with direct target-only supervision. Abstract-CoT-style representations were used during the development and teacher-data workflow, but reasoning traces are not output labels and the released model is instructed to return translation text only.
Training sources used for the releasable model include:
- Tatoeba bilingual sentence pairs distributed through OPUS under CC BY 2.0 FR.
- Project-authored technical terminology and correction templates under CC0 1.0.
AI Hub and DeepL samples were used only for held-out comparison/evaluation and were not included as training targets in this release.
Evaluation snapshot
- Correction strict checks: 12/12
- General core-meaning checks: 12/12
- Technical core-meaning checks: 12/12
- Empty output, trace leakage, and source echo: 0 in the recorded suites
- Technical reference similarity: 0.8779 for the selected adapter versus 0.8302 for the base candidate
These are small project evaluation suites and must not be interpreted as a comprehensive benchmark.
Checksums
105a9b734d32688c4f9c00bd16f88109dcfbf33d63da9645f11f2a9375dcbbb0c monarda-g4-e4b-base-q8_0.gguf
214fda53b2d83b30db930f9deb19838b4196f878a80c3226a2dfb45771f822b05 monarda-translation-correction-focus-f16.gguf
License and attribution
The base model is
Google Gemma 4 E4B IT, released under Apache License 2.0. This repository is released under Apache License 2.0. Tatoeba-derived training pairs require their original attribution under CC BY 2.0 FR.
llama.cpp is not included in this model repository. Liquid Translate distributes its runtime separately under the llama.cpp MIT license.
