How to Create a Reliable LLM Judge: The hh.ru Experience

Creating an LLM judge is easy. It’s much harder to ensure that its evaluations can be trusted. We learned this firsthand while developing a resume analysis feature for the hh.ru AI assistant.

The Trust Issue

It quickly became clear that a good LLM evaluator is a standalone product with its own evaluation criteria, datasets, quality metrics, and operational costs. Naive approaches, such as a simple prompt, don’t work: the evaluations end up being inconsistent and unreliable.

Judge Architecture

We designed an evaluation system that includes:

• Clear evaluation criteria (e.g., alignment of skills, experience, and education).
• Specialized datasets for training and validation.
• Quality metrics that allow us to track the accuracy of evaluations.

Each rubric has its own set of criteria, and the LLM evaluator is trained on annotated examples. This has allowed us to improve the consistency of evaluations and reduce the number of errors.

Conclusions

Key takeaway: The LLM judge requires the same attention to quality as the main product. You can’t just take a model and hope it will evaluate correctly. You need:

• Datasets annotated by experts.
• Metrics for continuous monitoring.
• Iterations to improve the model based on feedback.

This is the only way to build trust in automated evaluations.

Source: Habr