Pro-Bias: Self-Improving Human-Aligned Subject LLM Evals

Creating human-aligned LLM judges through enhanced evaluation frameworks and self-improving automated grading systems.

Comment

Pro-Bias: Self-Improving Human-Aligned Subjective LLM Evals

Inspiration

Inconsistent and misaligned subjective evaluations plague LLM development. We set out to create a framework that brings human-like judgment to AI evaluations.

What it does

  1. ComparisonGEval: Enhanced framework for consistent, human-aligned subjective LLM evaluations
  2. Synthetic datasets: Demonstrate ComparisonGEval's effectiveness across diverse tasks
  3. Example Evals: Showcase versatility in text casualization, conversation naming, and more
  4. Automated Essay Grading Agent: Iteratively improves rubrics to align with human raters

How we built it

  • Enhanced GEval with structured prompts and choice-based scoring
  • Generated synthetic datasets using Claude 3.5 Sonnet
  • Developed example evals for various subjective tasks
  • Created an agent that iteratively refines grading rubrics

Challenges we ran into

  • Ensuring consistency in subjective evaluations
  • Generating diverse, representative synthetic data
  • Aligning AI judgments with human raters

Accomplishments that we're proud of

  • Achieved substantial agreement with human raters on representative samples
  • Developed a versatile framework applicable to various subjective tasks
  • Created an iterative system that improves itself to match human judgment (!! HOLY GRAIL ALERT !!)

What we learned

  • The importance of structured prompts in subjective evaluations
  • Techniques for generating effective synthetic datasets
  • Strategies for aligning AI systems with human judgment

What's next

  • Expand to more complex subjective tasks
  • Integrate with popular LLM development workflows
  • Explore applications in educational technology and content moderation

Try it out

  1. Clone the repo
  2. Set up environment: Python 3.11, virtualenv, requirements
  3. Configure API keys: OpenAI/Sambanova, Anthropic
  4. Run evals: NUM_EXAMPLES=10 ./run_python.sh python evals/eval_make_text_more_casual.py
  5. Optimize essay rubrics with AI agent: ./run_python.sh python src/agents/essay_rubric_optimizer.py

View on GitHub

Built With

Share this project:

Updates