annotation ml fundamentals

Label Noise

Errors or inconsistencies in a dataset's annotations that reduce model training quality by corrupting the learning signal.

Label noise refers to errors, inconsistencies, or inaccuracies in the human-annotated ground truth data used to train a machine learning model. In the context of LLM alignment, label noise is poison to the training pipeline.

Label noise can originate from several sources. **Objective errors** occur when a rater simply makes a mistake, such as failing to spot a hallucinated fact or missing a syntax error in generated code. **Subjective noise** occurs when guidelines are ambiguous, causing two equally skilled raters to score the same response differently based on personal bias. **Adversarial noise** occurs when malicious or lazy workers intentionally provide garbage data to speed through tasks and maximize their hourly payout.

The impact of label noise is severe. If a model is fine-tuned on a dataset where 10% of the 'correct' responses contain hallucinations, the model will learn that hallucinating is a valid, rewarded behavior. During the RLHF phase, noisy preference data will cause the reward model to output chaotic, unpredictable scores, permanently degrading the final model's alignment.

AI labs combat label noise through relentless quality assurance: utilizing overlapping reviews (having multiple people rate the same prompt and taking the consensus), deploying honeypot questions, and using AI models to double-check human work for obvious errors.

What this means for trainers

Your primary imperative as an AI trainer is to produce zero-noise data. Skipping a fact-check because it takes too long, or guessing on a coding task because you aren't familiar with the language, introduces label noise. High-quality platforms will instantly flag and remove raters who inject noise into the training pipeline.