An expert-annotated dialogue bank for validating and calibrating LLM-based judges of multi-turn CBT-style therapy conversations. The benchmark accompanies the THERAPYGYM submission to the NeurIPS 2026 Evaluations & Datasets Track.

Anonymous release for double-blind review. Author identity will be revealed upon acceptance.

What It Is and What It Is Not

It is a calibration set for therapy-judge LLMs: 116 simulated patient–therapist dialogues, each rated… See the full description on the dataset page: https://huggingface.co/datasets/neurips-ed-2026-sub3717/therapyjudgebench.