A specialization set pairs real agent prompts with the API responses they are meant to reproduce. Capture the exact prompt and response, retain only well-formed outputs, and favor examples whose downstream execution confirmed they worked. section 12.2
Store each example with three fields:
- The prompt as input and API response as target. section 12.2
- A flag recording whether later agent execution confirmed the call; use it to filter or weight training examples. section 12.2
- If live traffic is too sparse, supplement it with teacher responses to generated inputs, recognizing that this is a weaker signal. section 12.2
{"input": "Book a table for two tomorrow at 8 pm",
"target": {"tool_name": "book_table",
"parameters": {"party_size": 2, "time": "20:00"}},
"confirmed": true}A sample record shows a prompt paired with structured tool arguments and a confirmation flag. section 12.2
Use the set to measure task-specific importance when pruning a model and to train the pruned model on the teacher’s hard completions during recovery. section 12.3
