Notes on Lecture by Sophia Young on Fine-Tuning Models

Introduction

• Speaker: Sophia Young, Lead Developer Relations at MOL
• Overview of the talk:
  • Overview of M models
  • Introduction of Fine-Tune API
  • Open-source fine-tuning codebase
  • Demos

Company Overview

• MOL is based in Paris with a team of over 50 people.
• Founded about a year ago.
  • September Last Year: Released first model, MR 7B.
  • December: Released Mixol (Time SB).
  • Commercial Model: MR Medium and API platform for model use.
  • February: Released MR Small and MR Large (flagship model with advanced capabilities).
  • April: Released open-source model 8 times 22B, best of its kind at the time.
  • Recent Release: Craw, specialized model trained on 80+ programming languages.

Model Offerings

• Three open-source models available for personal or commercial use.
• Two enterprise-grade models: MR Small and MR Large.
  • MR Large:
    • Supports multilingual function calling.
    • Specialized for retrieval-augmented generation (RAG).
  • Fine-tuning support available for MR Small and MR 7B.
• Emphasis on customization and user-specific needs.

Fine-Tuning Overview

• Fine-Tune API: Released to allow customization of models directly.
• Laura Fine-Tuning: Efficient and performant; analysis showed similar performance between Laura fine-tuning and full fine-tuning on MR 7B and MR Small.
• Comparison Results:
  • Laura Fine-Tuning: 0.9
  • Full Fine-Tuning: 0.91

Prompting vs Fine-Tuning

• Prompting:
  • Allows out-of-the-box functionality without data or training.
  • Easily updated for new workflows or prototyping.
• Fine-Tuning Advantages:
  • Often better performance than prompting.
  • Can work faster and more economically than lengthy prompts.
  • Better alignment with specific tasks due to focused training.

Demos

Demo Setup

• Installation: Ensure the latest version (0.4.0) of MOL API is installed.
• Using Fine-Tuned Models:
  • Example of generating abstracts from research paper titles.
  • Chatbot example using a medical dataset.
• Model Naming Convention:
  • Structure of model names shows the model it is fine-tuned on.

Case Studies

• Showcased various developer examples using the Fine-Tune API.
• Applications include:
  • Internet retrieval
  • Medical domain applications
  • Financial conversation assistants
  • Legal co-pilots

End-to-End Example

• Preparation Steps:
  • Install MOL AI and required packages.
  • Prepare and format the dataset for training.
  • Upload dataset to the server and define the model for fine-tuning.
  • Monitor training jobs and retrieve metrics.

Open Source Codebase

• Use the open-source codebase for fine-tuning MR 7B and other models.
• Example of downloading model and preparing data in Google Colab.
• Important to define configuration files for hyperparameters and paths.

Conclusion and Event Announcement

• Exciting news: Hosting a Fine-Tune Hackathon from today to June 30th.
• Encouragement to participate and showcase builds.

Questions and Answers

• Validation of data format is available through validation scripts.
• Support for further inquiries on fine-tuning processes.