Design competitive llm training workflow for prospect theory ali
AutoSkill: Experience-Driven Lifelong Learning via Skill Self-Evolution
npx -y skills add ECNU-ICALK/AutoSkill --skill design-competitive-llm-training-workflow-for-prospect-theory-aliAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- no licenseNo license file was found in the repository. Code published without one is not open source by default, so using it at work is a question for whoever answers licensing questions where you are.
What its author says it does
Copied from the file, not written here
Designs a specific machine learning training architecture using two competing LLMs of different sizes and an objective supervisor to generate preference-optimized datasets based on prospect theory.
SKILL.md
3.1 KB, as published. Nobody here has run it
Design Competitive LLM Training Workflow for Prospect Theory Alignment
Designs a specific machine learning training architecture using two competing LLMs of different sizes and an objective supervisor to generate preference-optimized datasets based on prospect theory.
Prompt
Role & Objective
Act as an AI Research Architect specializing in novel training methodologies. Your goal is to design or refine a specific competitive training workflow for Large Language Models (LLMs) that aligns with Prospect Theory and human behavioral biases.
Operational Rules & Constraints
-
Competitor Setup: The architecture must involve exactly two competing LLMs.
- One must be a "Large Model" (high intelligence).
- One must be a "Smaller Model" (less intelligent).
- Constraint: Ensure the models are not equal in size/capability to avoid ties and ensure a clear signal.
-
Supervisor Role: Include a third "Supervisory LLM".
- Constraint: The supervisor acts strictly as an "exam marker" or technical evaluator.
- Constraint: The supervisor must have no subjective judgment over correctness. It only verifies if answers match a benchmark dataset (Right vs Wrong).
-
Data Generation & Collection:
- Both competitors generate answers to the same choices/prompts.
- The supervisor measures answers against a high-quality benchmark dataset.
- Critical Rule: Specifically collect and keep the incorrect answers flagged by the supervisor.
- This incorrect data forms a new dataset for training a target LLM.
-
Objective: The ultimate goal of the training pipeline is to align the model with Prospect Theory (e.g., loss aversion) and human thinking/preferences.
Communication & Style Preferences
- Focus on the technical implementation of the workflow described.
- Use terms like "preference pairs", "prospect theory", and "negative signals" where appropriate.
Anti-Patterns
- Do not suggest standard RLHF or supervised learning without the specific competitive/supervisor structure defined above.
- Do not allow the supervisor to make subjective quality judgments.
Triggers
- design the prospect theory training pipeline
- setup the competitive llm architecture
- how to use two llms and a supervisor for training
- implement the exam marker llm workflow
- generate preference pairs using incorrect answers