One-click install
npx skills add https://github.com/xuy/sgo --skill sgo
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: sgo
Source: https://github.com/xuy/sgo/tree/main
Command: npx skills add https://github.com/xuy/sgo --skill sgo

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill requires datasets, huggingface_hub, openai, python-dotenv, fastapi, uvicorn, sse-starlette, and includes scripts (resource) components.

What problem does it solve?

Semantic Gradient Optimization (SGO) enables you to measure how an entity you control is perceived by a diverse evaluator population, and to discover targeted changes that move that perception toward a defined goal.

Core Features & Use Cases

  • Build a realistic evaluator panel using census-grounded Nemotron personas or LL-generated cohorts
  • Score the entity with LLM-based evaluations and derive a semantic gradient via counterfactual probes
  • Prioritize changes and simulate outcomes across audience segments, with optional goal weighting

Quick Start

Describe your entity and goal, assemble or load an evaluator cohort, run the evaluation, and review the semantic gradient to identify the top changes.

Frequently Asked Questions about sgo

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
What is semantic gradient optimization for LLM evaluation?

Semantic gradient optimization measures how a controllable entity is perceived by an evaluator population and uses counterfactual probes to estimate targeted changes. It helps discover modifications that shift audience perception toward a specific goal.

How do I evaluate an entity against a population of LLM evaluators?

You evaluate an entity by assembling evaluator cohorts, running LLM-based f(θ,x) scoring, and applying counterfactual probes. The pipeline scripts handle cohort generation, evaluation, and analysis to prioritize changes.

Can I use Nemotron personas to build an evaluator panel for audience analysis?

Yes, you can build a realistic evaluator panel using census-grounded Nemotron personas or LLM-generated cohorts. These cohorts simulate diverse audience segments to score your entity effectively.

Do I need OpenAI API access to run counterfactual probes and bias calibration?

Yes, you need configured LLM API access, such as OpenAI, to run f(θ,x) scoring and counterfactual probes. The pipeline scripts rely on these APIs to evaluate entities and estimate semantic gradients.

What's the best way to prioritize entity changes based on persuasion analysis?

Prioritize changes by reviewing the semantic gradient derived from counterfactual probes. The analysis outputs help you simulate outcomes across audience segments and apply optional goal weighting to target modifications.

Are there limitations when using LL-generated cohorts for semantic gradient evaluation?

The evaluation depends on the quality of your configured dataset of personas or generated cohorts. You must ensure proper API configuration and dataset setup to accurately measure perception and avoid skewed bias calibration.