Track 10: Training LLMs, Training Data
Deduplication and Decontamination
Keep test data out of training data.
Part of Train and Fine-Tune Models (25 of 45)
Course
Part of the free syllabus, not yet in a certificate course.
Udemy: coming soon
Tools
- Hugging Face Datasets (open-source, free)
Watch first
In this module
- Chat and Instruction Data Formats
- Build a 300-1,000 Example Dataset: Clean, Split, Licence-Check
- Synthetic Data With Guardrails
- Deduplication and Decontamination