Customising AI

[admin]

Multiple Solutions

The bedrock of Partition AI is for everyone to improve AI models independently - without Cloud Bottleneck.

1. Fine Tuning - expensive
2. Transfer Learning - cheaper
3. Retrieval Augment Generation - cheapest

Our focus in on Retrieval Augment Generation (RAG) but other methods are also supported.

More details on the differences between Transfer Learning and Fine Tuning:

• Understanding the Differences: Fine-Tuning vs. Transfer Learning - DEV Community
• https://www.dxtalks.com/blog/news-2/unlocking-llm-training-transfer-learning-vs-fine-tuning-explained-544

[admin]

Transfer Learning

Freeze all the pre-trained layers and only train the new layers added on top.

Some good introductory materials on transfer learning are listed here.

• What Is Transfer Learning? [Examples & Newbie-Friendly Guide]
• https://machinelearningmastery.com/transfer-learning-for-deep-learning/
• Some videos:
  https://www.youtube.com/watch?v=yofjFQddwHE
  https://www.youtube.com/watch?v=FQM13HkEfBk

Fine Tuning

Unfreeze some of the pre-trained layers and allow them to be updated during training.

There are MANY ways to make Fine Tuning faster:

Guide to fine-tuning LLMs using PEFT and LoRa techniques

RAG

RAG vs LoRA

• RAG vs. LoRA for Efficient Language Model Tuning | by A. J. S. Raja Ravi Varman | Medium