A dataset is a structured collection of data used to find patterns or train a computational model. It can contain labeled examples, such as images paired with verified descriptions, or unlabeled material. For generative AI, very large and varied datasets supply real-world examples for training—because apparently the internet had to become homework first. section 1.8
Datasets also drive analysis. Before software can reliably answer questions about a large spreadsheet or transcript, its contents must be made machine-readable: the unglamorous but vital work called data wrangling. An LLM-based workflow can load documents, build an index, and answer targeted questions, such as which regions have the highest average household net wealth. section 8.1
A practical dataset workflow has a few moving parts:
- Choose data suited to the question: a transcript for summarizing a call, a spreadsheet for business analysis, or reviews for sentiment analysis. section 8.1 section 8.4
- Keep useful labels when they exist. Prelabeled reviews make it possible to compare model-generated sentiment with known sentiment values. section 8.4
- Query or process the data with clear constraints, then inspect the results rather than treating a model response as magic confetti. section 8.1 section 8.4
- Match training data to a specialized model’s domain. Legal or medical specialization calls for a dataset rich in legal or medical content, plus substantial hardware for fine-tuning. section 9.5 section 9.8
Checking sentiment labels
A CSV dataset of 1,000 X messages about companies includes each comment and a sentiment label. Python loads the file, sends every comment to GPT-3 and GPT-3.5 with a positive-or-negative prompt, and stores the predicted labels beside the existing ones. Counting the combinations exposes agreement, mistakes, ambiguous answers, and operational trouble such as rate-limit failures. A smaller sample may be necessary when the API refuses to play along. section 8.4
| Company |
Sentiment |
Comment |
| Microsoft |
Negative |
@Microsoft Why do I pay for WORD when it functions so poorly on my @SamsungUS Chromebook? |
| MaddenNFL |
Positive |
Thank you @EAMaddenNFL!! |
Inspect a small labeled-review table to see the kind of paired company, sentiment, and comment data used for automated sentiment testing. section 8.4
Creating a dataset for model fine-tuning is a separate, demanding job. It requires enough relevant content to train effectively and is paired with significant hardware needs; the detailed build process is deliberately left for another adventure. section 9.5 section 9.8
