Annotation

Process data published

Also known as: Data Labeling, Labeling, Tagging

Definition

The process of adding labels, tags, or metadata to data to make it suitable for supervised learning or evaluation. For LLMs, annotation includes labeling text for classification, rating response quality for RLHF, marking entities for NER, and creating instruction-response pairs for fine-tuning.

What this is NOT

  • Not synthetic data generation (annotation is human labeling)
  • Not data collection (annotation adds labels to existing data)
  • Not model training (annotation prepares data for training)

Alternative Interpretations

Different communities use this term differently:

llm-practitioners

Human labeling of data for LLM training and evaluation: rating response quality, choosing preferred outputs for RLHF, labeling safety issues, or creating gold-standard evaluation sets.

Sources: RLHF annotation practices, Scale AI, Labelbox documentation, Data labeling literature

Examples

  • Rating two model responses as 'A is better' for DPO training
  • Labeling responses as 'helpful', 'harmless', 'honest'
  • Creating instruction-response pairs for fine-tuning
  • Annotating text with entity labels for NER

Counterexamples

Things that might seem like Annotation but are not:

  • Generating data with an LLM (that's synthetic)
  • Scraping web data (no labeling involved)
  • Model self-evaluation (not human annotation)

Relations

  • overlapsWith human-feedback (Annotation produces human feedback data)
  • overlapsWith rlhf (RLHF requires preference annotations)
  • overlapsWith dataset (Annotation produces labeled datasets)
  • overlapsWith benchmark (Benchmarks require annotated evaluation data)

See Also

Implementations

Tools and frameworks that implement this concept: