Tracking Large Language Models (LLM) with MLflow : A Complete Guide

As Large Language Models (LLMs) grow in complexity and scale, tracking their performance, experiments, and deployments becomes increasingly challenging. This is where MLflow comes in – providing a comprehensive platform for managing the entire lifecycle of machine learning models, including LLMs.

In this in-depth guide, we’ll explore how to leverage MLflow for tracking, evaluating, and deploying LLMs. We’ll cover everything from setting up your environment to advanced evaluation techniques, with plenty of code examples and best practices along the way.

Functionality of MLflow in Large Language Models (LLMs)

MLflow has become a pivotal tool in the machine learning and data science community, especially for managing the lifecycle of machine learning models. When it comes to Large Language Models (LLMs), MLflow offers a robust suite of tools that significantly streamline the process of developing, tracking, evaluating, and deploying these models. Here’s an overview of how MLflow functions within the LLM space and the benefits it provides to engineers and data scientists.

Tracking and Managing LLM Interactions

MLflow’s LLM tracking system is an enhancement of its existing tracking capabilities, tailored to the unique needs of LLMs. It allows for comprehensive tracking of model interactions, including the following key aspects:

• Parameters: Logging key-value pairs that detail the input parameters for the LLM, such as model-specific parameters like top_k and temperature. This provides context and configuration for each run, ensuring that all aspects of the model’s configuration are captured.
• Metrics: Quantitative measures that provide insights into the performance and accuracy of the LLM. These can be updated dynamically as the run progresses, offering real-time or post-process insights.
• Predictions: Capturing the inputs sent to the LLM and the corresponding outputs, which are stored as artifacts in a structured format for easy retrieval and analysis.
• Artifacts: Beyond predictions, MLflow can store various output files such as visualizations, serialized models, and structured data files, allowing for detailed documentation and analysis of the model’s performance.

This structured approach ensures that all interactions with the LLM are meticulously recorded, providing a comprehensive lineage and quality tracking for text-generating models​.

Evaluation of LLMs

Evaluating LLMs presents unique challenges due to their generative nature and the lack of a single ground truth. MLflow simplifies this with specialized evaluation tools designed for LLMs. Key features include:

• Versatile Model Evaluation: Supports evaluating various types of LLMs, whether it’s an MLflow pyfunc model, a URI pointing to a registered MLflow model, or any Python callable representing your model.
• Comprehensive Metrics: Offers a range of metrics tailored for LLM evaluation, including both SaaS model-dependent metrics (e.g., answer relevance) and function-based metrics (e.g., ROUGE, Flesch Kincaid).
• Predefined Metric Collections: Depending on the use case, such as question-answering or text-summarization, MLflow provides predefined metrics to simplify the evaluation process.
• Custom Metric Creation: Allows users to define and implement custom metrics to suit specific evaluation needs, enhancing the flexibility and depth of model evaluation.
• Evaluation with Static Datasets: Enables evaluation of static datasets without specifying a model, which is useful for quick assessments without rerunning model inference.

Deployment and Integration

MLflow also supports seamless deployment and integration of LLMs:

• MLflow Deployments Server: Acts as a unified interface for interacting with multiple LLM providers. It simplifies integrations, manages credentials securely, and offers a consistent API experience. This server supports a range of foundational models from popular SaaS vendors as well as self-hosted models.
• Unified Endpoint: Facilitates easy switching between providers without code changes, minimizing downtime and enhancing flexibility.
• Integrated Results View: Provides comprehensive evaluation results, which can be accessed directly in the code or through the MLflow UI for detailed analysis.

MLflow is a comprehensive suite of tools and integrations makes it an invaluable asset for engineers and data scientists working with advanced NLP models.

Setting Up Your Environment

Before we dive into tracking LLMs with MLflow, let’s set up our development environment. We’ll need to install MLflow and several other key libraries:

pip install mlflow>=2.8.1
pip install openai
pip install chromadb==0.4.15
pip install langchain==0.0.348
pip install tiktoken
pip install 'mlflow[genai]'
pip install databricks-sdk --upgrade

After installation, it’s a good practice to restart your Python environment to ensure all libraries are properly loaded. In a Jupyter notebook, you can use:

import mlflow
import chromadb
print(f"MLflow version: {mlflow.__version__}")
print(f"ChromaDB version: {chromadb.__version__}")

This will confirm the versions of key libraries we’ll be using.

Understanding MLflow’s LLM Tracking Capabilities

MLflow’s LLM tracking system builds upon its existing tracking capabilities, adding features specifically designed for the unique aspects of LLMs. Let’s break down the key components:

Runs and Experiments

In MLflow, a “run” represents a single execution of your model code, while an “experiment” is a collection of related runs. For LLMs, a run might represent a single query or a batch of prompts processed by the model.

Key Tracking Components

1. Parameters: These are input configurations for your LLM, such as temperature, top_k, or max_tokens. You can log these using mlflow.log_param() or mlflow.log_params().
2. Metrics: Quantitative measures of your LLM’s performance, like accuracy, latency, or custom scores. Use mlflow.log_metric() or mlflow.log_metrics() to track these.
3. Predictions: For LLMs, it’s crucial to log both the input prompts and the model’s outputs. MLflow stores these as artifacts in CSV format using mlflow.log_table().
4. Artifacts: Any additional files or data related to your LLM run, such as model checkpoints, visualizations, or dataset samples. Use mlflow.log_artifact() to store these.

Let’s look at a basic example of logging an LLM run:

This example demonstrates logging parameters, metrics, and the input/output as a table artifact.

import mlflow
import openai
def query_llm(prompt, max_tokens=100):
    response = openai.Completion.create(
        engine="text-davinci-002",
        prompt=prompt,
        max_tokens=max_tokens
    )
    return response.choices[0].text.strip()
with mlflow.start_run():
    prompt = "Explain the concept of machine learning in simple terms."
    
    # Log parameters
    mlflow.log_param("model", "text-davinci-002")
    mlflow.log_param("max_tokens", 100)
    
    # Query the LLM and log the result
    result = query_llm(prompt)
    mlflow.log_metric("response_length", len(result))
    
    # Log the prompt and response
    mlflow.log_table("prompt_responses", {"prompt": [prompt], "response": [result]})
    
    print(f"Response: {result}")

Deploying LLMs with MLflow

MLflow provides powerful capabilities for deploying LLMs, making it easier to serve your models in production environments. Let’s explore how to deploy an LLM using MLflow’s deployment features.

Creating an Endpoint

First, we’ll create an endpoint for our LLM using MLflow’s deployment client:

import mlflow
from mlflow.deployments import get_deploy_client
# Initialize the deployment client
client = get_deploy_client("databricks")
# Define the endpoint configuration
endpoint_name = "llm-endpoint"
endpoint_config = {
    "served_entities": [{
        "name": "gpt-model",
        "external_model": {
            "name": "gpt-3.5-turbo",
            "provider": "openai",
            "task": "llm/v1/completions",
            "openai_config": {
                "openai_api_type": "azure",
                "openai_api_key": "{{secrets/scope/openai_api_key}}",
                "openai_api_base": "{{secrets/scope/openai_api_base}}",
                "openai_deployment_name": "gpt-35-turbo",
                "openai_api_version": "2023-05-15",
            },
        },
    }],
}
# Create the endpoint
client.create_endpoint(name=endpoint_name, config=endpoint_config)

This code sets up an endpoint for a GPT-3.5-turbo model using Azure OpenAI. Note the use of Databricks secrets for secure API key management.

Testing the Endpoint

Once the endpoint is created, we can test it:

response = client.predict(
endpoint=endpoint_name,
inputs={"prompt": "Explain the concept of neural networks briefly.","max_tokens": 100,},)
print(response)
This will send a prompt to our deployed model and return the generated response.
Evaluating LLMs with MLflow
Evaluation is crucial for understanding the performance and behavior of your LLMs. MLflow provides comprehensive tools for evaluating LLMs, including both built-in and custom metrics.
Preparing Your LLM for Evaluation
To evaluate your LLM with mlflow.evaluate(), your model needs to be in one of these forms:
An mlflow.pyfunc.PyFuncModel instance or a URI pointing to a logged MLflow model.
A Python function that takes string inputs and outputs a single string.
An MLflow Deployments endpoint URI.
Set model=None and include model outputs in the evaluation data.
Let's look at an example using a logged MLflow model:
import mlflow
import openai
with mlflow.start_run():
    system_prompt = "Answer the following question concisely."
    logged_model_info = mlflow.openai.log_model(
        model="gpt-3.5-turbo",
        task=openai.chat.completions,
        artifact_path="model",
        messages=[
            {"role": "system", "content": system_prompt},
            {"role": "user", "content": "{question}"},
        ],
    )
# Prepare evaluation data
eval_data = pd.DataFrame({
    "question": ["What is machine learning?", "Explain neural networks."],
    "ground_truth": [
        "Machine learning is a subset of AI that enables systems to learn and improve from experience without explicit programming.",
        "Neural networks are computing systems inspired by biological neural networks, consisting of interconnected nodes that process and transmit information."
    ]
})
# Evaluate the model
results = mlflow.evaluate(
    logged_model_info.model_uri,
    eval_data,
    targets="ground_truth",
    model_type="question-answering",
)
print(f"Evaluation metrics: {results.metrics}")

This example logs an OpenAI model, prepares evaluation data, and then evaluates the model using MLflow's built-in metrics for question-answering tasks.

Custom Evaluation Metrics
MLflow allows you to define custom metrics for LLM evaluation. Here's an example of creating a custom metric for evaluating the professionalism of responses:

from mlflow.metrics.genai import EvaluationExample, make_genai_metric
professionalism = make_genai_metric(
    name="professionalism",
    definition="Measure of formal and appropriate communication style.",
    grading_prompt=(
        "Score the professionalism of the answer on a scale of 0-4:n"
        "0: Extremely casual or inappropriaten"
        "1: Casual but respectfuln"
        "2: Moderately formaln"
        "3: Professional and appropriaten"
        "4: Highly formal and expertly crafted"
    ),
    examples=[
        EvaluationExample(
            input="What is MLflow?",
            output="MLflow is like your friendly neighborhood toolkit for managing ML projects. It's super cool!",
            score=1,
            justification="The response is casual and uses informal language."
        ),
        EvaluationExample(
            input="What is MLflow?",
            output="MLflow is an open-source platform for the machine learning lifecycle, including experimentation, reproducibility, and deployment.",
            score=4,
            justification="The response is formal, concise, and professionally worded."
        )
    ],
    model="openai:/gpt-3.5-turbo-16k",
    parameters={"temperature": 0.0},
    aggregations=["mean", "variance"],
    greater_is_better=True,
)
# Use the custom metric in evaluation
results = mlflow.evaluate(
    logged_model_info.model_uri,
    eval_data,
    targets="ground_truth",
    model_type="question-answering",
    extra_metrics=[professionalism]
)
print(f"Professionalism score: {results.metrics['professionalism_mean']}")

This custom metric uses GPT-3.5-turbo to score the professionalism of responses, demonstrating how you can leverage LLMs themselves for evaluation.

Advanced LLM Evaluation Techniques
As LLMs become more sophisticated, so do the techniques for evaluating them. Let's explore some advanced evaluation methods using MLflow.

Retrieval-Augmented Generation (RAG) Evaluation
RAG systems combine the power of retrieval-based and generative models. Evaluating RAG systems requires assessing both the retrieval and generation components. Here's how you can set up a RAG system and evaluate it using MLflow:

from langchain.document_loaders import WebBaseLoader
from langchain.text_splitter import CharacterTextSplitter
from langchain.embeddings import OpenAIEmbeddings
from langchain.vectorstores import Chroma
from langchain.chains import RetrievalQA
from langchain.llms import OpenAI
# Load and preprocess documents
loader = WebBaseLoader(["https://mlflow.org/docs/latest/index.html"])
documents = loader.load()
text_splitter = CharacterTextSplitter(chunk_size=1000, chunk_overlap=0)
texts = text_splitter.split_documents(documents)
# Create vector store
embeddings = OpenAIEmbeddings()
vectorstore = Chroma.from_documents(texts, embeddings)
# Create RAG chain
llm = OpenAI(temperature=0)
qa_chain = RetrievalQA.from_chain_type(
    llm=llm,
    chain_type="stuff",
    retriever=vectorstore.as_retriever(),
    return_source_documents=True
)
# Evaluation function
def evaluate_rag(question):
    result = qa_chain({"query": question})
    return result["result"], [doc.page_content for doc in result["source_documents"]]
# Prepare evaluation data
eval_questions = [
    "What is MLflow?",
    "How does MLflow handle experiment tracking?",
    "What are the main components of MLflow?"
]
# Evaluate using MLflow
with mlflow.start_run():
    for question in eval_questions:
        answer, sources = evaluate_rag(question)
        
        mlflow.log_param(f"question", question)
        mlflow.log_metric("num_sources", len(sources))
        mlflow.log_text(answer, f"answer_{question}.txt")
        
        for i, source in enumerate(sources):
            mlflow.log_text(source, f"source_{question}_{i}.txt")
    # Log custom metrics
    mlflow.log_metric("avg_sources_per_question", sum(len(evaluate_rag(q)[1]) for q in eval_questions) / len(eval_questions))