Back to Agno

Persona-Driven Generation

cookbook/data_labeling/_24_persona_driven_generation/README.md

2.8.04.2 KB
Original Source

Persona-Driven Generation

PersonaHub-style conditioning: a typed persona (occupation, expertise level, communication style, current concern) steers the generator, so the same domain yields different registers, concerns, and vocabulary - a novice trucker and an M&A attorney do not ask the same retirement question. Every row carries its full persona as provenance, and the diversity gain is measured with counted lexical metrics, not asserted.

Files

  • basic.py - a persona agent invents 6 distinct personas in one call; a prompt agent writes 2 questions per persona about a fixed domain (personal finance). Rows carry the full persona.
  • math_problems.py - 6 hand-written personas condition unit-rate multiplication word problems. The problem shape is pinned (exactly two whole numbers in the text, answer = their product), so a pure-Python check re-extracts the numbers and re-derives every gold answer from the problem text; rows whose stated answer fails the check are dropped and counted. The verified output feeds ../_21_rejection_sampling/.
  • diversity_report.py - measures what conditioning buys: 8 unconditioned prompts vs 8 persona-conditioned prompts about the same domain, compared on distinct-1, distinct-2, and mean pairwise Jaccard distance (all pure stdlib). No JSONL - the printed report is the artifact.

Rows are written to data/generated/ (gitignored - run the scripts to regenerate). Rows from a real run:

json
{"prompt": "My trucking fleet doesn't offer a 401(k) match, so I need to set up my own retirement account. I don't want some broker eating up my hard-earned money with hidden charges. Where can I open a simple, low-fee IRA where the rules are easy to understand and I won't get ripped off by fine print?", "persona": {"occupation": "Commercial Truck Driver", "expertise_level": "novice", "communication_style": "plainspoken, direct, and skeptical of financial jargon", "current_concern": "Finding a reliable, low-fee individual retirement account since the trucking fleet employer does not offer a 401(k) matching program."}}
{"problem": "With feed prices climbing, I need to closely calculate our daily rations. Each cow in my milking herd requires 6 pounds of the new energy grain mix per day. If I currently have 74 cows to feed, how many pounds of this grain mix do I need for the whole herd each day?", "answer": 444, "persona_occupation": "dairy farmer"}
{"problem": "Hurry, I need to restock the supply carts before my night shift gets crazy. I have 15 carts to fill. Each cart gets exactly 6 sterile suture kits. How many total suture kits must I gather?", "answer": 90, "persona_occupation": "emergency room nurse"}

Measured result, honestly

The register and topical spread of conditioned prompts is visibly wider, but at N=8 the lexical metrics only partly capture it. In the logged run, mean pairwise Jaccard distance rose (0.886 -> 0.908), distinct-2 moved within run-to-run noise, and distinct-1 consistently FELL (0.642 -> 0.562)

  • because conditioned prompts average roughly 3x more tokens (18.5 -> 59.4) and distinct-n is length-sensitive: longer prompts repeat more function words regardless of topical spread. The report prints mean tokens per prompt alongside the metrics so this confound stays visible. Treat distinct-n comparisons across pools of different lengths with suspicion; if the direction matters to you, hold length constant or use a length-insensitive measure.

When to use

When you need coverage of voices, not just tasks: user simulation, question mining for a fixed domain, or surface-form variety over a fixed skill (as in math_problems.py, where personas vary the story while the arithmetic stays checkable).

Run

bash
python cookbook/data_labeling/_24_persona_driven_generation/basic.py
python cookbook/data_labeling/_24_persona_driven_generation/math_problems.py
python cookbook/data_labeling/_24_persona_driven_generation/diversity_report.py

Requires GOOGLE_API_KEY.