Fine-tuning Language Models for Structured Responses with QLoRa

Опубликовано: 09 Август 2023
на канале: Trelis Research

17,500

572

I cover fine-tuning of language models to return *structured responses*, e.g. to return function calls, json objects or arrays. Lecture notes here: https://colab.research.google.com/dri...

Fine-tuning for tone or style? • Fine-tuning Llama 2 for Tone or Style

Basic Training Google Colab Notebook (FREE)
Access the Google Colab script here: https://colab.research.google.com/dri...

ADVANCED Training Notebook for Structured Responses (PAID)
Includes a prompt loss-mask and stop token for improved performance.
Learn more: https://trelis.com/function-calling/

Advanced Fine-tuning Repo Access - incl. 5+ advanced notebooks
Learn more here: https://trelis.com/advanced-fine-tuni...
1. Fine-tuning for structured responses
2. Supervised fine-tuning (best for training "chat" models)
3. Unsupervised fine-tuning (best for training "base" models)
4. Embeddings generation and usage (alternative to fine-tuning)

Function Calling Dataset
Function calling dataset: https://huggingface.co/datasets/Treli...

Out-of-the-box Llama 2 with Function Calling
Llama 70b, 34b (code llama), 13b, 7b: https://huggingface.co/Trelis/Llama-2...

0:00:00 Understanding Model Size
0:03:56 Quantization
0:09:09 Loading and Setting Up a Training Notebook
0:15:26 Data Setup and Selection
0:15:52 Training Process
0:19:31 Inference and Prediction
0:23:17 Saving and Push the Model to the Hub
0:25:59 ADVANCED Fine-tuning and Attention tutorial