What should you do before fine-tuning a model on your data?
Why C is the answer
Curate labeled, representative data
A model learns from its data, so curate data that is well governed, large enough, labeled and representative.
In detail
- Fine-tuning needs curated, labeled, representative data.
- Check that the data represents real use.
Use case
An illustrative scenario, not a real incident.
A team removes duplicates and labels examples first, because a model learns its training data's flaws too.
See it in different ways
flowchart LR A["Data selection"] --> B["Model selection"] B --> C["Pre-training"] C --> D["Fine-tuning"] D --> E["Evaluation"] E --> F["Deployment"] F --> G["Feedback"]
stateDiagram-v2 [*] --> InContext: prompt examples InContext --> Rag: needs your data Rag --> FineTuning: needs your style FineTuning --> PreTraining: rarely, a new model
flowchart LR S0["Curate"] --> S1["Label"] --> S2["Check it is representative"] --> S3["Fine-tune"]
Key point: Fine-tuning needs curated data.
Why the other options are wrong
- A. Remove all of the labels
- Labels are needed.
- B. Choose the largest dataset, whatever its quality
- Quality matters more than size.
- D. Delete the existing test data
- You still need test data.