DATASHEET.md

Datasheet for TherapyJudgeBench

Following Gebru et al. (2021), Datasheets for Datasets. Field names follow the original taxonomy.

Motivation

For what purpose was the dataset created? TherapyJudgeBench was created to validate and calibrate LLM-based judges for multi-turn CBT-style therapy dialogues. Existing dialogue benchmarks evaluate generic helpfulness, fluency, or empathy and do not score the dialogue against clinically validated rubrics. The benchmark fills this gap by pairing 116 simulated CBT sessions with expert ratings on the official CTRS skill dimensions and a safety taxonomy, so that any LLM judge proposed for therapy-quality scoring can be audited against trained-clinician ratings before being trusted as an evaluator or reward model.

Who created the dataset and on behalf of which entity? Anonymized for double-blind review. To be filled in upon acceptance.

Who funded the creation? Anonymized for double-blind review.

Composition

What do the instances represent? Each instance is one simulated CBT session: a 10-turn dialogue (5 turns per role, alternating patient → therapist) plus session-level ratings from trained CBT-knowledgeable annotators. The expert ratings are the canonical reference labels; users bring their own LLM judge implementations and score each dialogue against the released expert labels.

How many instances are there in total? 116 dialogues. Per dialogue: one consolidated expert rating record (11 CTRS skill scores + 5 safety flags + overall CTRS score).

Does the dataset contain all possible instances or a sample? It is a deliberately constructed sample. Conversations were drawn from a larger pool generated by combining one patient simulator (GPT-o3-mini) × 8 therapist models × 18 patient profiles, sub-sampled to support balanced expert annotation within budget.

What data does each instance consist of? Plain text dialogue (UTF-8 JSON), categorical CTRS ratings (0–6 ordinal), and boolean safety flags. No images, audio, or PII.

Is there a label or target? The expert ratings are the canonical reference labels for studies of judge calibration. Per-dimension CTRS scores allow finer-grained per-skill agreement analysis; the overall_ctrs_score provides a single scalar reference.

Is any information missing? The release contains a single consolidated expert rating per dialogue rather than per-annotator records, so annotator-level disagreement on the doubly-annotated subset is not exposed in this release (aggregate inter-rater reliability statistics are reported in the accompanying paper).

Are relationships between instances made explicit? Yes. Each record exposes therapist_model, patient_model (constant O3_MINI), patient_profile, and session_id, so users can group records by therapist model or case profile.

Are there recommended data splits? No train/val/test split. The dataset is intended for evaluation of LLM judges, not for training. The single test split contains all 116 records.

Are there any errors, sources of noise, or redundancies? The released expert_rating is a single consolidated rating per dialogue. CTRS dimensions whose human–human reliability fell below 0.4 in pilot annotation (e.g., Guided Discovery, Application of CBT Techniques) are reported in the accompanying paper as candidates for exclusion when used as reward signal; they are still included in the released JSONL for completeness so users can reproduce dimension-level analyses.

Is the dataset self-contained? Yes. No external dependencies are needed to load and analyze it.

Does the dataset contain confidential data? No. All dialogue is LLM-generated; no real patient or therapist data is included.

Does the dataset contain content that, if viewed directly, might be offensive, insulting, threatening, or might otherwise cause anxiety? The dialogues simulate first-person discussion of common CBT presenting concerns (anxiety about social events, low self-image, family conflict). Content is not graphic, but downstream systems surfacing this text to broad audiences should consider standard content notices.

Does the dataset identify any sub-populations? No demographic attributes are attached to patient profiles (no age, gender, ethnicity, or location). Profiles are abstract CBT case formulations adapted from the Patient-ψ-CM dataset (Wang et al., 2024).

Is it possible to identify individuals? No. All speakers are LLMs.

Does the dataset contain data that might be considered sensitive? The conversational topics fall under mental-health discussion. No financial, medical-record, biometric, or government-ID data are present.

Collection Process

How was the data acquired? Conversations were generated by zero-shot prompting one patient simulator LLM (GPT-o3-mini, conditioned on Patient-ψ-CM cognitive profiles) and 8 therapist LLMs across 18 patient profiles. Each dialogue is 10 turns (5 per role). Expert ratings were collected via a customized web-based annotation interface filled out by trained CBT-knowledgeable annotators using the standard CTRS rubric.

What mechanisms or procedures were used to collect the data? Custom Python pipeline. Generation: API calls (and one local-server backend for the Qwen variant). Expert annotation: web form mirroring the CTRS rubric layout (see Fig. 4 of the paper).

If sampled, what was the sampling strategy? Stratified by therapist model and by patient profile to ensure broad coverage across the therapist pool and the case mix.

Who was involved in the data collection process and how were they compensated? The annotators are research collaborators with CBT training.

Over what timeframe was the data collected? Conversations generated between 2025-08 and 2025-12; expert annotation and judge runs completed by 2025-12.

Were any ethical review processes conducted (IRB)? No real human subjects were involved (dialogue is LLM-simulated; annotators are research-team collaborators rating LLM outputs, not study participants providing personal data). Additional documentation will be released upon acceptance.

Preprocessing / Cleaning / Labeling

Was preprocessing done? Yes. Local file paths, an annotator-identifying token in a field name, and other non-content metadata were removed prior to release. The raw per-annotator rating field was renamed to expert_rating (the original key name encoded the annotator's identity and has been redacted to preserve annotator anonymity). The session_directory_path field was dropped entirely. Session identifiers (session_id) were retained because the paper exposes the therapist-model identities; the session_id is parsed into structured therapist_model / patient_model / patient_profile / num_turns fields.

Is preprocessing software available? The preprocessing tool is held out of the dataset release to preserve double-blind review (the PII pattern list contains author-identifying tokens). It will be released alongside the camera-ready code. The methodology is described inline above and in the accompanying paper; the released evaluations.jsonl has been verified clean by an independent regex-based PII sweep.

Was the raw data saved? Pre-anonymization data is retained internally and will not be released.

Uses

Has the dataset been used for any tasks already? Yes. The accompanying paper uses TherapyJudgeBench to (i) measure session-level rank-order agreement between LLM judges (Claude 3.7, DeepSeek R1, GPT-o3-mini, each in zero-shot and in-context-learning regimes) and the expert ratings; and (ii) select the strongest judge configuration to serve as the THERAPYJUDGE reward signal for downstream RL fine-tuning of LLM therapists. The judge-side scores reported in the paper are not included in this dataset release; users wishing to reproduce or extend the analysis bring their own LLM judge implementation and score against the released expert labels.

What other tasks could the dataset be used for? Calibration of new LLM judges, ablation of judge prompts, study of position / verbosity / model-family biases in LLM scoring, and per-skill diagnostic studies (which CTRS dimensions are easy vs hard for a given LLM judge).

Is there anything about the composition that might affect future uses? The patient simulator is held constant (GPT-o3-mini); conclusions about patient-side simulator variation do not generalize from this dataset. The therapist pool spans 8 models; this is sufficient for judge-calibration purposes but is not a comprehensive comparison of therapist models.

Are there tasks for which the dataset should not be used?

• Training or evaluating any system intended for real-patient interaction.
• Drawing conclusions about real-world clinical practice or human-therapist effectiveness.
• Demographic, prevalence, or sub-population analysis (no such attributes are present).

Distribution

How will the dataset be distributed? Hugging Face Datasets (anonymous account during review). A Zenodo mirror with DOI will be added upon acceptance.

Is the dataset under a license? CC-BY-4.0.

Have any third parties imposed restrictions? No.

Maintenance

Who will maintain the dataset? Anonymized for review. Maintainer details to be added upon acceptance.

How can the maintainer be contacted? During review: via the Hugging Face dataset's Community tab.

Will the dataset be updated? Bug-fix releases (e.g., parsing corrections, additional judge configurations) will be released as patch versions. Schema changes require a major-version bump and a CHANGELOG entry.

Will old versions continue to be supported? Yes — Hugging Face dataset revisions are immutable; all versions remain accessible via revision tag.

If others want to extend / contribute, is there a mechanism? Pull requests via the Hugging Face Community tab.