gyung/lfm2-1.2b-koen-mt-v8-rl-10k-merged
The gyung/lfm2-1.2b-koen-mt-v8-rl-10k-merged model is a 1.2 billion parameter Korean-English machine translation model developed by Gyung, based on LiquidAI's LFM2-1.2B. It achieves state-of-the-art performance for its size in both Korean-to-English and English-to-Korean translation, significantly outperforming larger 4B models like Gemma-3 on the CHrF++ benchmark. This model is specifically fine-tuned using Group Relative Policy Optimization (GRPO) with a reward model incorporating COMET and CHrF++, excelling in maintaining consistent honorifics and natural sentence structures in translation.
Loading preview...
Overview
The gyung/lfm2-1.2b-koen-mt-v8-rl-10k-merged model is a 1.2 billion parameter Korean-English machine translation model developed by Gyung. It is built upon the LiquidAI LFM2-1.2B base model and has been fine-tuned using Group Relative Policy Optimization (GRPO) with a reward model combining COMET and CHrF++. This model achieves state-of-the-art performance for its size, notably surpassing 4B parameter models like Gemma-3 on the Flores-200 benchmark (CHrF++ score of 34.61 vs 32.83).
Key Capabilities
- Superior Translation Quality: Achieves SOTA performance for its size in both Korean-to-English and English-to-Korean translation, with only 400 GRPO training steps.
- Honorific Consistency: Demonstrates perfect consistency in applying honorifics (e.g., "~합니다") across all tested samples.
- Natural Sentence Structure: Capable of translating complex sentences into natural-sounding structures.
- Context-Aware Translation: Adapts word choices based on context (e.g., translating "While" appropriately).
- Efficient Performance: A 1.2B model that outperforms larger 4B models, making it efficient for deployment.
Good For
- Korean-English Machine Translation: Ideal for applications requiring high-quality, context-aware, and honorific-consistent translation between Korean and English.
- Resource-Constrained Environments: Its small size (1.2B parameters) combined with strong performance makes it suitable for deployment on devices with limited computational resources, including mobile or edge devices (especially with recommended GGUF quantization).
- Further Research and Development: The fp32 merged version is recommended for additional fine-tuning or research due to its full parameter access.
Known Limitations
- Proper Noun Hallucination: Inherits biases from the base model, leading to incorrect translations of some proper nouns (e.g., "George W. Bush" to "조지 워싱턴").
- Specific Technical Terminology: May exhibit errors in translating certain specialized scientific terms.