Manual dataset building creates a fine-tuning set by writing every prompt–completion pair yourself. It offers close control over a model’s behavior for a given input, but requires substantial time because every example is curated individually. section 7.4.1
Teach a title pattern
A pair can ask for “a title for the general public about topic X” and answer “X revealed to you!” After fine-tuning, a relativity prompt may yield a similarly framed title. Build separate, specific datasets for audiences, scenarios, or output types such as titles, annotations, and commentary. section 7.4.1
Use a practical workflow:
- Curate samples that cover all intended cases; at least 10 samples are required. section 7.4.1
- Train the model, then give identical prompts to the original and fine-tuned models. section 7.4.1
- Adopt the fine-tuned model only when its outputs improve on the original. section 7.4.1
Table 7.2 The samples based on the information to convey
|
Information to Convey
|
Prompt
|
Completion
a
|
|---|---|---|
|
Parts of a whole
|
Percentage of participants in the conference by type (researchers 5%, students 30%, professors 65%)
|
More professors participated in the conference (65%). Researchers were not interested in the event (5%).
|
|
Comparison among entities
|
Comparison between red (80%), yellow, and green performance
|
Compared to yellow and green, red experienced an improvement in performance of 80%.
|
|
Trend
|
Sales changed in the last 12 months (–12%), due to fewer subscribers and video views.
|
Over the last 12 months, sales decreased by 12% based on various metrics, including subscribers and video views.
|
|
Outcomes of a survey or a questionnaire
|
Questionnaire: 3 positive answers, 7 negative answers
|
Three out of 10 people answered the questionnaire with a positive answer.
|
|
Distribution
|
Sales of product A (+30%) and product B over the last 12 months
|
Compared to product B, sales of product A have increased by 30% over the last 12 months.
|
|
Spatial information
|
Sales in North Europe (+23%) compared to South Europe
|
Compared to South Europe, sales in North Europe increased by 23%.
|
|
Relationship
|
The sales trend line from 1990 to 2020 increased by 120%.
|
Between 1990 and 2020, sales increased by 120%.
|
|
Comparison among entities
|
Top ingredients for our recipe: sugar and salt
|
The chosen ingredients for our recipe are sugar and salt.
|
|
Comparison among entities
|
Comparison between gold (30), silver (20), and bronze (40)
|
Bronze beats silver and gold with 40.
|
|
Distribution
|
Distribution of household chores (cooking 35%, cleaning 30%, laundry 20%, and yard work 15%)
|
Cooking takes up the most significant portion at 35%. Cleaning follows at 30%, while laundry and yard work account for 20% and 15%, respectively.
|
| a. The word completion may be confusing, but it is used by the OpenAI API. Completion refers to the output produced by the model. | ||
Sample pairs show how prompts about parts of a whole, comparisons, trends, surveys, and distributions can be matched with concise data commentary. section 7.4.1
