Instruction data teaches a language model to transform source content according to user requests. Each example pairs an instruction and original draft with the desired revision, making editing behavior explicit. section 8.2.3
Build a useful instruction dataset by:
- Collecting drafts and real revision patterns across content types such as whitepapers, blog posts, and case studies. section 8.2.3
- Pairing every instruction with its original and revised text so the requested transformation is observable. section 8.2.3
- Balancing simple edits, such as shortening, with complex changes, such as combining sections for clarity. section 8.2.3
- Reserving coordination and data-generation resources because filling gaps takes more effort than expected. section 8.2.3
Table 8.3 Data sample for instruction fine-tuning
| Input |
Target output |
|
| Instruction |
Original text |
Revised text |
| “Simplify the language for a wider audience.” |
“This workflow orchestration mechanism enables seamless process automation across ERP systems.” |
“This system makes it easy to automate workflows in ERP software.” |
| “Add an example to illustrate the argument.” |
“The solution improves efficiency.” |
“The solution improves efficiency. For example, it reduces order processing time by 30% for one of our clients.” |
| “Restructure to highlight key points earlier.” |
“The software includes advanced features. The introduction outlines key benefits for users.” |
“The software offers key benefits such as ease of use and scalability. Advanced features are detailed below.” |
Review paired examples showing simplification, evidence addition, and restructuring from original text to target output. section 8.2.3
