Skip to main content
An efficient bidirectional translation system powered by Liquid AI’s LFM2 1.2B model fine-tuned for Korean-English translation. This project demonstrates how domain-specific fine-tuning can achieve superior performance compared to models 3x larger.

​
What it does

This example provides a high-quality translation CLI with:
  • Automatic language detection - Intelligently detects input language and translates accordingly
  • Superior performance - CHrF++ 34.61 / BLEU 13.21 on Flores-200 benchmark, outperforming Google’s Gemma-3 4B and Alibaba’s Qwen3 4B
  • Efficient inference - Runs on modest hardware with merged adapters for speed
  • Easy-to-use CLI - Simple command-line interface powered by Fire
This project was built and released by Kiwoong Yeom with the support of Maxime Labone. View the original announcement on LinkedIn.

​
Performance benchmarks

Tested on Flores-200 benchmark (1,012 samples):
ModelParametersCHrF++BLEU
LFM2-KoEn-v8-rl1.2B34.6113.21
Gemma-3-4B4B32.8311.36
Qwen3-4B4B25.627.46
The 1.2B parameter model outperforms models 3x larger, demonstrating that specialized training matters more than raw parameter count.

​
Quick start

1

Clone the repository

git clone https://github.com/Liquid4All/cookbook.git
cd cookbook/examples/lfm2-english-to-korean
2

Install dependencies

uv sync
3

Run translation example

uv run python main.py --text "$(cat linkedin_post.txt)" --max-new-tokens 1024

​
Understanding the architecture

The system uses a two-stage training approach:
  1. Supervised fine-tuning (SFT): 280K high-quality Korean-English parallel datasets establish the translation foundation
  2. Reinforcement learning (RL): GRPO optimization with 10K additional samples refines translation quality

​
Model components

  • Base model: gyung/lfm2-1.2b-koen-mt-v6.4-merged - SFT fine-tuned LFM2 1.2B
  • Adapter: gyung/lfm2-1.2b-koen-mt-v8-rl-10k-adapter - LoRA adapter trained with GRPO
  • Automatic detection: Regular expression pattern matching for Korean text (Hangul syllables, Jamo)
The system automatically detects the input language and applies the appropriate translation direction, supporting both English→Korean and Korean→English translation.

​
CLI usage

uv run python main.py [OPTIONS]

Options:
  --text TEXT                    Text to translate (required)
  --model-name TEXT             Base model name (default: gyung/lfm2-1.2b-koen-mt-v6.4-merged)
  --adapter-name TEXT           Adapter name (default: gyung/lfm2-1.2b-koen-mt-v8-rl-10k-adapter)
  --max-new-tokens INTEGER     Maximum tokens to generate (default: 256)
  --temperature FLOAT          Sampling temperature (default: 0.3)
  --min-p FLOAT               Minimum probability threshold (default: 0.15)
  --repetition-penalty FLOAT   Repetition penalty (default: 1.05)

​
Example usage

uv run python main.py --text "Hello, how are you today?"

​
Further improvements

Next steps for enhanced performance and efficiency:
  • Speed optimization with quantization techniques (GGUF, AWQ, GPTQ)
  • llama.cpp integration for faster CPU inference
  • Full parameter RL training with expanded compute resources
  • Length normalization removal based on recent Qwen team findings
  • Extended dataset training with 200K SFT + 25K RL samples

​
Performance optimization

The current implementation uses adapter merging for faster inference. Future improvements include:
  • Quantized model variants for resource-constrained environments
  • Streaming inference for real-time translation
  • Batch processing for large document translation

​
Source code

View the complete source code on GitHub.

Build docs developers (and LLMs) love

Get started for freeTalk to us