Learn how to build a QLoRA fine-tuning dataset for language models in this comprehensive video tutorial. Explore various Reddit dataset options, including torrent files, Archive.org, and BigQuery. Follow step-by-step instructions on exporting BigQuery Reddit data, decompressing gzip archives, and recombining archives for target subreddits. Discover the proper data structure, build training samples, and save them to a database. Create customized training JSON files and dive into QLoRA training and results. Gain valuable insights into dataset building for fine-tuning language models through practical demonstrations and explanations.
Building an LLM Fine-Tuning Dataset - From Reddit Comments to QLoRA Training
Overview
Syllabus
- Introduction to Dataset building for fine-tuning.
- The Reddit dataset options Torrent, Archive.org, BigQuery
- Exporting BigQuery Reddit and some other data
- Decompressing all of the gzip archives
- Re-combining the archives for target subreddits
- How to structure the data
- Building training samples and saving to database
- Creating customized training json files
- QLoRA training and results
Taught by
sentdex
