👉 Japanese Article of this repository

This repository provides a practical guide and working example for integrating MLflow Prompt Engineering UI with non-officially supported Japanese LLMs, specifically from the LLM-jp project.

MLflow Prompt Engineering UI is an experimental feature allowing users to test and optimize prompts in a no-code UI. It supports OpenAI and other hosted models, but local models like those from LLM-jp require manual setup.

This repository shows how to:

• Use MLflow Model Serving to serve local LLM-jp models via pyfunc.PythonModel
• Expose the local models through MLflow AI Gateway
• Run prompt optimization experiments via Prompt Engineering UI

1. serve local models via pyfunc.PythonModel

python wrap_pyfunc.py

2. MLflow Tracing and Streamlit UI

mlflow ui --backend-store-uri sqlite:///mlruns/mlflow.db --port 5000

streamlit run app.py --server.fileWatcherType none

3. AI gateway and prompt enginnering

export MLFLOW_DEPLOYMENTS_TARGET="http://127.0.0.1:7000"
export OPENAI_API_KEY="" # if you use OpenAI Model for LLM-as-a-judge

mlflow models serve -m ./saved_model/ --no-conda --port 5001

mlflow gateway start --config-path config.yaml --port 7000
mlflow server --port 5000

• Tracking Experiments with a Local Database

The MLflow client can interface with a SQLAlchemy-compatible database (e.g., SQLite, PostgreSQL, MySQL) for the backend. Saving metadata to a database allows you cleaner management of your experiment data while skipping the effort of setting up a server.

graph TD;
    A[Hugging Face Hub Model Card] -->|Convert to| B[MLflow PyFunc Model];
    B -->|Save Locally| C[Local MLflow Model Storage];
    C -->|Serve with MLflow| D[MLflow Model Serving];
    D -->|Expose via AI Gateway| E[MLflow AI Gateway Endpoint];
    E -->|Use in| F[MLflow Prompt Engineering UI];

Japanese prompt input may cause encoding issues in Prompt Engineering UI.
Workaround: use ASCII-only prompts or sanitize input manually.

• LLM-jp:
  A Cross-organizational Project for the Research and
  Development of Fully Open Japanese LLMs
• LLM-jp: 日本語に強い大規模言語モデルの研究開発を行う
  組織横断プロジェクト
• MLflow Tracing for LLM Observability
  
• LLM 勉強会
• Hugging Face LLM-jp
• Prompt Engineering UI (Experimental)
  
• MLflow AI Gateway (Experimental)
  

• LLM-jp-3 Fine-tuned Models
  • llm-jp/llm-jp-3-150m-instruct2
  • llm-jp/llm-jp-3-150m-instruct3
  • llm-jp/llm-jp-3-440m-instruct2
  • llm-jp/llm-jp-3-440m-instruct3
  • llm-jp/llm-jp-3-980m-instruct2
  • llm-jp/llm-jp-3-980m-instruct3
  • llm-jp/llm-jp-3-1.8b
  • llm-jp/llm-jp-3-1.8b-instruct
  • llm-jp/llm-jp-3-1.8b-instruct2
  • llm-jp/llm-jp-3-1.8b-instruct3
  • llm-jp/llm-jp-3-3.7b
  • llm-jp/llm-jp-3-3.7b-instruct
  • llm-jp/llm-jp-3-3.7b-instruct2
  • llm-jp/llm-jp-3-3.7b-instruct3
  • llm-jp/llm-jp-3-7.2b-instruct
  • llm-jp/llm-jp-3-7.2b-instruct2
  • llm-jp/llm-jp-3-7.2b-instruct