DiaCBT: Multi-Session CBT Dialogue Corpus

• DiaCBT is a long-periodic dialogue corpus that provides synthetic, multi-session CBT interactions with detailed strategy annotations and Cognitive Conceptualization Diagrams.
• It is created through a rigorous pipeline combining expert-annotated transcripts, CCD-guided dialogue generation, and stringent human evaluation to ensure CBT fidelity.
• Empirical findings show that models fine-tuned on DiaCBT achieve enhanced CBT-specific performance, demonstrating improved simulated therapist skills and longitudinal treatment consistency.

DiaCBT is a long-periodic dialogue corpus for psychological counseling based on Cognitive Behavioral Therapy (CBT). It is a synthetic but CBT-faithful resource designed to train and evaluate psychotherapy dialogue agents over multiple sessions rather than within isolated, short interactions. The corpus contains 108 complete CBT cases, each spanning five sessions, for a total of 540 sessions, with multi-turn therapist–client dialogues, fine-grained CBT strategy annotations for therapist segments, and Cognitive Conceptualization Diagrams (CCDs) that ground each case in an explicit cognitive model of the client (Zhou et al., 3 Sep 2025).

1. Origins, Motivation, and Research Scope

DiaCBT was created in response to two constraints identified in LLM-based psychotherapy research. First, real multi-session psychotherapy transcripts are scarce and usually not shareable because of privacy and sensitivity. Existing open resources rely heavily on brief, single-session, or even single-turn interactions, often resolve a client’s problem in one short dialogue, and frequently emphasize depression or generic emotional support rather than a full CBT treatment course. Second, general-purpose LLMs may produce empathetic responses, but they are not trained on longitudinal, strategy-annotated CBT dialogues and therefore do not reliably emulate therapist behavior guided by CBT case formulations (Zhou et al., 3 Sep 2025).

The corpus is intended to model the full CBT process across multiple sessions, including early assessment, automatic thought work, belief restructuring, behavior change, and relapse prevention. It also uses CCDs to simulate realistic clients with consistent cognitive profiles and to guide utterance generation across diverse scenarios. In parallel, it provides strategy-level annotations for therapist behavior through 14 CBT techniques, enabling models to learn both when and how to use CBT strategies. The associated research questions are explicitly developmental and evaluative: whether a model fine-tuned on DiaCBT can better emulate expert CBT therapists, whether the corpus improves CBT-specific skills beyond generic counseling, and whether a rigorous evaluation framework can reflect CBT clinical standards.

The phrase “long-periodic” is central to the dataset’s scope. In this setting, it denotes multiple sessions per case and a therapy trajectory that unfolds over time, with strategy distributions shifting across sessions rather than repeating a single counseling template. This design places DiaCBT closer to psychotherapy process modeling than to general supportive dialogue generation.

2. Corpus Design and Cognitive Conceptualization Diagrams

The final dataset contains 108 cases, 540 sessions, and 2613 strategy-coherent segments. At the case level, the average number of utterances is 264.87 and the average number of tokens is 6253.20. At the session level, the average number of utterances is 52.97 and the average number of tokens per utterance is approximately 23.8 (Zhou et al., 3 Sep 2025).

Component	Value
Cases	108
Sessions	540
Segments	2613
Sessions per case	5
Avg. utterances per case	264.87
Avg. utterances per session	52.97

The corpus is grounded in the CBT notion of cognitive conceptualization. In Beck’s framework, a Cognitive Conceptualization Diagram is a structured representation of how core beliefs, intermediate beliefs, automatic thoughts, emotions, and behaviors interrelate. The full CCD structure includes eight components—Relevant History, Core Beliefs, Intermediate Beliefs, Coping Strategies, Situation, Automatic Thoughts, Emotions, and Behaviors—but DiaCBT uses a six-component subset: Core Beliefs, Intermediate Beliefs, Situation, Automatic Thoughts, Emotions, and Behaviors. These components are instantiated for each synthetic client using CBT-related datasets such as C2D2 and PatternReframe as seed information (Zhou et al., 3 Sep 2025).

A typical CCD in DiaCBT includes statements such as core beliefs (“I am out of control. I am undesirable, unwanted.”), intermediate beliefs (“I'm just not very good at handling stress and I have poor self-control, which is why I need to not put myself in stressful situations.”), a situation (“I just got promoted recently, but I'm afraid I'm not up to the task.”), automatic thoughts, emotions, and behaviors. The CCD is not merely metadata. It defines who the client is cognitively, guides LLM-based client role-play, and conditions therapist strategy selection and formulation. Therapists are therefore modeled as working through the CBT hierarchy from automatic thoughts toward intermediate and core beliefs.

This structure gives DiaCBT a formal case-level representation that many prior counseling corpora lack. A plausible implication is that the dataset supports research on longitudinal therapeutic consistency, because client behavior can be judged against an explicit cognitive formulation rather than only against local prompt content.

3. Data Creation Pipeline and Expert Filtering

DiaCBT was constructed through a four-step pipeline: case annotation from expert sources, cognitive conceptualization via CCDs, CCD-guided dialogue generation, and expert evaluation and filtering (Zhou et al., 3 Sep 2025).

The first step used public CBT transcripts from APA PsycTherapy and a CBT textbook, “Dispelling the Fog of Belief.” From 53 original transcripts, the authors derived 33 sessions and then selected 6 complete consultation processes as high-quality exemplars. CBT-trained psychology graduate students annotated therapist segments with one of the 14 strategy labels, following a detailed guideline and a training and competency test. These annotations were then used as few-shot exemplars for generation.

The second step produced CCDs. Using CBT-related datasets such as C2D2, described as a Chinese cognitive distortion dataset with 7 categories, and PatternReframe, the authors used LLMs including GPT-4o to construct six-component CCDs:

$$\text{CCD}=\{\text{Core Beliefs},\text{Intermediate Beliefs},\text{Situation},\text{Automatic Thoughts},\text{Emotions},\text{Behaviors}\}.$$

These conceptualizations were later checked by experts, and more than 80% of ratings for each component were “very” or “extremely” accurate.

The third step used script mode for dialogue generation rather than two LLMs conversing autonomously. Prior work cited in the paper, CACTUS, is reported to have shown that script mode yields more coherent and natural dialogues than two-agent mode. For each case, the pipeline selected a CCD and a CBT session plan, provided the CCD together with few-shot annotated CBT segments and a scripted instruction prompt to GPT-4o-mini, and asked the model to generate the entire multi-session dialogue. The instructions required adherence to CBT session flow, use of Socratic and open-ended questions, and a facilitative stance that helps the client discover solutions rather than being told what to think.

The fourth step imposed human screening. Five trained annotators refined raw dialogues, and experts sampled and audited each annotator’s work. The evaluation criteria were Correctness, Reasonableness, and Situation Diversity. Dialogues were sent back for revision if standards were not met, and stiff or simplistic therapist questions were rephrased into more natural clinical language. Of approximately 3600 raw segments, 2613 were retained, producing a 72.58% retention rate. The high rejection rate is explicitly interpreted by the authors as evidence that raw LLM psychotherapy ability is insufficient without structured prompting and expert filtering.

4. CBT Operationalization, Strategy Taxonomy, and Longitudinal Structure

DiaCBT operationalizes CBT through a single-label strategy taxonomy of 14 therapist strategies: Information Gathering, Setting the Agenda, Weekly Review, Defining Therapeutic Objectives, Psychoeducation, Working with Automatic Thoughts, Motivational Enhancement, Working with Intermediate and Core Beliefs, Behavioral Techniques, Relapse Prevention, Homework Assignments, Requesting Feedback, Summarization, and Other (Zhou et al., 3 Sep 2025).

These labels are defined functionally rather than descriptively. Information Gathering covers exploration of basic data, history, and current conditions. Setting the Agenda is collaborative session focusing. Weekly Review tracks prior sessions and progress. Defining Therapeutic Objectives converts presenting problems into concrete therapy goals. Psychoeducation explains disorders, CBT rationale, and mechanisms. Working with Automatic Thoughts addresses situation-linked thoughts through eliciting, examining, and testing them. Working with Intermediate and Core Beliefs surfaces and modifies deeper beliefs. Behavioral Techniques include behavior experiments, exposure, and activity scheduling. Relapse Prevention addresses maintenance and setbacks. Homework Assignments design and review out-of-session tasks. Requesting Feedback asks for client feedback on sessions. Summarization recaps key points.

The segment counts show a pronounced concentration in core CBT intervention categories. Working with Automatic Thoughts accounts for 438 segments and 10,962 utterances; Information Gathering for 303 segments and 7,565 utterances; Working with Intermediate and Core Beliefs for 192 segments and 5,398 utterances; Behavioral Techniques for 195 segments and 4,136 utterances; and Homework Assignments for 168 segments and 3,096 utterances. This distribution mirrors the paper’s claim that repeated work on thoughts and beliefs is central to CBT practice.

Longitudinal structure is explicit. Each case forms a five-session sequence. Session 1 emphasizes assessment, relationship building, agenda setting, and early thought work. Sessions 2 and 3 deepen work on automatic thoughts, begin behavioral experiments, and refine goals. Sessions 4 and 5 shift toward intermediate and core beliefs, relapse prevention, consolidation, and feedback. The strategy frequencies vary systematically by session, with early sessions showing more Information Gathering and Working with Automatic Thoughts, and later sessions showing more Working with Intermediate and Core Beliefs and longer, deeper interventions. This organization makes DiaCBT a longitudinal corpus rather than a collection of independent counseling episodes.

The CBT theory encoded in the corpus is also explicit. Situations do not directly cause emotions; interpretations in the form of automatic thoughts do. Automatic thoughts, intermediate beliefs, and core beliefs are treated as distinct levels of cognitive structure. DiaCBT’s CCD formalization is therefore both a data schema and a theory-grounded representation of psychotherapy process.

5. Model Training and Evaluation Framework

To demonstrate the dataset’s utility, the authors fine-tuned Qwen2.5-7B-Instruct with LoRA. The reported configuration uses LoRA rank 64, $\alpha = 16$, AdamW, learning rate $1\mathrm{e}{-4}$, batch size 32, 3 epochs, and a single NVIDIA A800 GPU. Training is framed as joint generation of therapist strategy and therapist utterance conditioned on the client utterance, the dialogue history, and an instruction template (Zhou et al., 3 Sep 2025):

$$\min_{\theta} \sum_{t=1}^{T} - \log p_{\theta} \left( s_t, u_t^d \mid I_{task}, h_{t-1}, u_t^c \right).$$

Here, $u_t^d$ is the therapist utterance at turn $t$, $s_t$ is the CBT strategy label at turn $t$, $u_t^c$ is the client utterance at turn $t$, $\alpha = 16$0 is the dialogue history, and $\alpha = 16$1 denotes model parameters.

The evaluation framework measures both therapist competence and simulated client change. For client simulation, the authors selected 140 cases from C2D2, with 20 per distortion type, built CCDs for those cases, and used GPT-4o as the AI client conditioned on the CCD and instructed to stay in character. Sessions continued until the client output an end token such as “goodbye” or a maximum turn limit was reached. Expert ratings again found that more than 80% of CCD-component judgments were “very” or “extremely” accurate.

Automatic evaluation includes Average Turn (AT), Success Rate (SR), PANAS-based affect change, and CTRS-like counseling skill ratings. AT measures the average number of conversation turns per session. SR measures goal completion effectiveness and uses a third LLM grader, $\alpha = 16$2, to evaluate each dialogue state as “worse,” “same,” “better,” or “solved,” with the scalar mapping

$\alpha = 16$3

To reduce variance, the outputs are averaged over $\alpha = 16$4 samples:

$\alpha = 16$5

If $\alpha = 16$6, the session is treated as goal-completed. PANAS is used to score positive and negative affect before and after each session, yielding Positive $\alpha = 16$7 and Negative $\alpha = 16$8 metrics.

The counseling skill evaluation adapts the Cognitive Therapy Rating Scale to LLM evaluation. General skills are Understanding, Interpersonal Effectiveness, and Collaboration. CBT-specific skills are Guided Discovery, Focus on key cognitions/behaviors, and Strategy. Each is scored from 0 to 6, and the sum gives an aggregate score. Human evaluation sampled 140 dialogues and compared the DiaCBT model against PsyChat and CAMEL on Relevance, CBT Style Measure, and Helpfulness. Inter-rater agreement was reported as Fleiss’ Kappa = 0.685 with $\alpha = 16$9.

6. Empirical Findings, Position in the Literature, and Limitations

The fine-tuned DiaCBT model outperformed several baselines on multiple automatic and human evaluations. In automatic evaluation, the model achieved the highest Average Turn at 12.05, the highest Success Rate at 77.14%, and the highest PANAS Positive change at 1.675. Its PANAS Negative change was -1.021, described as slightly better than several baselines but not the best in reduction. In CTRS-like ratings, the model obtained the highest total sum score at 28.75, with particularly strong CBT-specific performance: Guided Discovery 5.03 and Strategy 4.42 were both the best results reported, while Focus was 4.31, second only to CAMEL’s 4.38 (Zhou et al., 3 Sep 2025).

Question-style analysis further characterizes the counseling behavior learned from DiaCBT. The fine-tuned model produced shorter therapist utterances than the standard model, with an average therapist utterance length of 31.13 tokens versus 61.15, while also producing more questions per dialogue, 9.04 versus 4.46, and more in-depth questions, 7.41 versus 2.75. The reported interpretation is that a standard LLM behaves more like an “advice giver,” whereas the DiaCBT-trained model interacts more like a CBT therapist through shorter, more interactive exchanges and a stronger emphasis on exploring clients’ thoughts in depth.

Human evaluation also favored the DiaCBT model. Against CAMEL, it won 88 dialogues and lost 34 overall; against PsyChat, it won 106 and lost 24. On CBT Style Measure, it recorded 76 wins and 53 losses versus CAMEL, and 92 wins and 17 losses versus PsyChat. On Helpfulness, it recorded 56 wins and 24 losses versus CAMEL, and 96 wins and 28 losses versus PsyChat. Relevance was the one axis on which CAMEL was reported as slightly better, with 51 wins and 89 losses from DiaCBT’s perspective, a result the paper associates with CAMEL’s strong but often very closed questioning.

Within the broader literature represented here, DiaCBT occupies a distinct role. CoCoA is a proof-of-concept CBT counseling agent built on top of an LLM, with a counseling-specific memory system focused on cognitive distortions and a dynamic prompting framework that plans which CBT technique and which stage to use before generating each response (Lee et al., 2024). DiaCBT, by contrast, is a corpus and benchmarking framework for long-periodic counseling, with CCD-grounded client simulation and explicit strategy annotations. A different line of work models CBT sessions as control-affine dynamical systems and extracts local dynamic modes from short conversational windows to predict therapist competence from transcripts, treating therapist utterances as control inputs and client utterances as observations (Ardulov et al., 2022). This places DiaCBT alongside generation-oriented and evaluation-oriented research rather than as a direct replacement for either.

The paper also positions DiaCBT against existing datasets. It states that DiaCBT is the only dataset among the compared corpora that is simultaneously explicitly CBT-based, multi-turn, full-session, strategy annotated, and CCD-guided. This uniqueness follows from the conjunction of longitudinal structure, explicit case conceptualization, and therapist-segment labeling.

Several limitations are explicit. Sessions remain shorter than real-life CBT sessions of about 45 minutes, and five sessions do not capture the full range of longer treatments. The corpus is synthetic and expert-filtered rather than composed of real client transcripts. The fine-tuned model shows slight strategy bias, with some strategies such as psychoeducation under-used relative to expert transcripts. The data sources combine English CBT materials and Chinese cognitive distortion datasets, so generalization across languages and cultures is not guaranteed. On safety, the authors state that no human feedback or RLHF was used for safety tuning, some responses may be harmful if used directly with vulnerable people, the model is not a replacement for real therapy, and any deployment should clearly disclose that responses are AI-generated and only informational.

Potential applications follow directly from the dataset design: training CBT-style chatbots, benchmarking counseling models with CCD clients and AT/SR/PANAS/CTRS metrics, therapist training with simulated patients, and research on strategy selection and long-horizon dialog planning. Future work identified in the paper includes longer and more realistic multi-session interactions, broader psychotherapeutic techniques beyond CBT, more nuanced annotations such as client reactions and symptom trajectories, addressing strategy bias, and possibly incorporating real-world data under strict privacy safeguards.