A dataset is an organized collection of records or examples assembled for analysis, model development, retrieval, or evaluation. In generative AI workflows, it may supply external knowledge, demonstrate a fine-tuning task, or provide test inputs for measuring model behavior. section 7.2 section 9.3 section 12.3.7
Use datasets in several complementary parts of the GenAI stack:
- Update an external RAG dataset as information changes, avoiding model retraining whenever new data is added. section 7.2
- Prepare task examples for fine-tuning, such as teaching a model to return Microsoft Teams-compatible emoji syntax. section 9.3
- Run model evaluations from CSV or JSONL test data, or enter evaluation records manually through Azure AI Studio. section 12.3.7
EmojiBot fine-tuning dataset
A fine-tuning dataset can encode the desired EmojiBot behavior: given a task involving Microsoft Teams syntax, the model learns to respond with emojis represented by parenthesized text such as (dog). The dataset therefore expresses both the task and the required output convention. section 9.3
Figure 12.1 Azure AI Studio evaluations
Use the Azure AI Studio evaluation setup to inspect how a test dataset supports evaluation runs, metrics, custom flows, and batch execution. section 12.3.7
