Critical CV-NLP Data Errors and How to Fix Them with Galileo
summary
Claim: 80/20 performance split — most gains come from fixing data, not model tinkering. Demonstrates concrete error types: mislabeled examples, class overlap, data imbalance, and drift, and specific detection methods: loss variance, model certainty, similarity search and clustering on embeddings for duplicates and drift. Shows a working toolchain: dq.auto trains a Hugging Face model and feeds results to the Galileo console, surfaces mislabeled/empty clusters and lets teams re-label or rebalance successfully.
AI-written summary · about
chapters (13)
- 0:00Critical data errors in ML
- 1:09Mislabeled data harms models
- 3:19Causes of mislabeled data
- 4:45Detect and fix mislabeled data
- 7:12Class overlap in datasets
- 8:59Solutions for class overlap
- 11:53Data imbalance impact
- 13:39Detect duplicates and outliers
- 16:29Fix imbalance with embeddings
- 18:59What is data drift
- 20:22Detecting and addressing drift
- 23:13Summary of critical errors
- 24:39Galileo demo
related talks
37mOptimal on Paper, Broken in Reality
01h 08mMastering Visual AI with Vision-Language Models & Advanced Evaluation Techniques
33mKeys to Data-Centric AI: Asking AI to Improve its own Dataset.
17mAnnotating Log Data For Machine Learning
35mImproving Credibility of ML Solutions
01h 33mUnderstanding and diagnosing your machine-learning models
48mCommon Data Science Mistakes
44mKeep it Clean: Why Bad Data Ruins Projects and How to Fix it