Lecture Notes: Overview of M Models and Fine-Tuning API

Introduction

• Speaker: Sophia Young, Lead Developer Relations at Mol.
• Excited to discuss M models, recently released Fine Tune API, and open-source fine-tuning codebase.

Company Overview

• Mrol: Founded in September last year, based in Paris with a team of over 50.
• Key Model Releases:
  • Mr 7B (Sept 2022)
  • Mixol and Mr Medium (Dec 2022)
  • Mr Small and Mr Large (Feb 2023)
    • Mr Large: Flagship model with advanced reasoning and multilingual capabilities.
  • 8x 22b (April 2023)
  • Craw: Specialized model trained on 80+ languages focusing on code generation.

Model Offerings

• Open Source Models: Three models available for personal/commercial use.
• Enterprise Models:
  • Mr Small: Fine-tuning support.
  • Mr Large: Advanced functionalities, multilingual, and function calling capabilities.
• Emphasis on customization and fine-tuning.

Fine-Tuning Overview

• Fine-Tuning Code Base: Released to allow users to fine-tune open-source models.
• Fine Tun API: Launched to customize models directly.
• Technology Used: Laura Fine-Tuning (efficient and performant).
  • Performance comparison:
    • Mr 7B (Laura FT): 0.9
    • Full Fine-Tuning: 0.91

Prompting vs. Fine-Tuning

• Prompting:
  • No data/training required; works out of the box.
  • Good for prototyping and quick updates.
• Fine-Tuning:
  • Can outperform larger models for specific use cases.
  • More aligned with tasks of interest; learns new facts/information.

Demos

1. Fine-Tune Model Demo:
   • Example: Input titles to generate research paper abstracts.
   • Showcasing a model fine-tuned on title-abstract pairs.
2. Medical Chatbot Example:
   • Trained on medical datasets to answer queries.
3. Data Generation:
   • Generating data using a larger model like Mr Large for fine-tuning purposes.
4. Real-World Use Cases:
   • Startups using fine-tuned models for various sectors (medical, finance, legal).

Using Fine-Tuning API

• Installation: Install latest Mol API (0.4.0).
• Data Preparation:
  • Use parquet files; ensure data sizes are within limits (training: <512MB, evaluation: <1MB).
• Job Creation:
  • Define model and hyperparameters, create fine-tuning jobs, and monitor progress through metrics.

End-to-End Example

• Codebase Setup:
  • Clone MR Fine Tune repo and install required packages.
• Define configurations, paths, and hyperparameters for fine-tuning.
• Start training and monitor the checkpoints for inference.

Exciting News

• Hackathon Announcement:
  • Hosting a fine-tuning hackathon from today until June 30th.
  • Participants can submit ideas through a Google form.

Conclusion

• Encouragement to explore the fine-tuning capabilities and participate in the hackathon.
• Thank you for attending!