RodolfoRizzi/MyMentorLLM-dataset

Dataset Card for MyMentorLLM

This dataset accompanies the MyMentorLLM paper, arXiv:2607.25667, which describes the simulation environment, generation procedure, experimental design and validation analyses. If you use this dataset, please cite the paper (see Citation Information).

Dataset Summary

MyMentorLLM is a multimodal voice- and text-based deliberate-practice environment used to generate 2,100 simulated Cognitive Behavioural Therapy (CBT) training sessions, with 65,100 dialogue turns and 263.4 hours of speech.

Each session includes three LLM roles:

• a patient with major depressive disorder (MDD), generalised anxiety disorder (GAD) or borderline personality disorder (BPD), informed by the vignettes in DSM-5-TR Clinical Cases (Barnhill, 2023);
• a therapist-in-training, who conducts the session;
• a mentor, acting as an expert supervisor, who scores the trainee on the 11-item Cognitive Therapy Rating Scale (CTRS) and runs a structured debrief on diagnosis and symptoms.

Sessions	2,100
Dialogue turns	65,100
Audio clips	27,900 FLAC, 20.3 GiB, in Parquet and as separate files
Language	Italian
Models	Gemini 3.1 Live, Gemini 3.5 Flash (supervision only), Gemma 4 (12B, E2B), Qwen 3.5 9B, Qwen 3.6 35B

Analysis code is available at github.com/RodolfoRizzi/MyMentorLLM.

Dataset Structure

Data Instances

Component	Configuration	File	Rows	Each row
Therapy turns	turns	data/therapy_turns.parquet	65,100	One turn of the therapy dialogue
Mentor sessions	sessions	data/mentor_sessions.parquet	2,100	One session with its mentor evaluation
Audio turns	audio_turns	data/audio_turns/*.parquet	27,900	One spoken turn, with its audio and text
Audio index		metadata.csv	27,900	One FLAC file under audio/

sessions and turns join on session_uid or session_id. audio_turns and metadata.csv join turns on session_id and turn_index. Sessions and turns are kept in separate tables because they are different units of analysis. Parquet files are canonical; .xlsx copies are also provided.

A condition is model_patient_therapist × modality: 7 conditions of 300 sessions each.

from datasets import load_dataset

sessions = load_dataset("RodolfoRizzi/MyMentorLLM-dataset", "sessions")
turns = load_dataset("RodolfoRizzi/MyMentorLLM-dataset", "turns")
audio = load_dataset("RodolfoRizzi/MyMentorLLM-dataset", "audio_turns")  

The prompts that generated the sessions are in prompts/.

audio_turns is sharded by model: hf download RodolfoRizzi/MyMentorLLM-dataset --repo-type dataset --include "data/audio_turns/gemma12b-*" fetches one model only.

Data Fields

Complete information about field types and the model identifiers are in docs/dataset_schema.md.

Session fields

Common Columns shared across files:

Column	Type	Files	Meaning
session_uid	string	all files	Stable public identifier, <model>_<modality>_<disorder>_<NNN>, e.g. gemma12b_text_MDD_042.
session_id	string	all files	Original run identifier.
model_patient_therapist	string	all files	Model used for both the patient and trainee therapist roles
patient_case	string	all files	Patient's disorder: BPD, GAD or MDD
modality	string	therapy_turns, mentor_sessions	audio (spoken conversation) or text (typed)
thinking	int	therapy_turns, mentor_sessions	Model reasoning enabled (1) or not (0)
model_mentor	string	mentor_sessions	Model used as mentor

model_mentor differs from model_patient_therapist only in the Gemini condition. The live speech model cannot return structured output, so gemini-3.5-flash was the mentor in those sessions.

therapy_turns.parquet: 65,100 × 11

The session fields, plus:

Column	Type	Meaning
turn_index	int	Turn order in the conversation, 0–30
timestamp	string	Turn time from the start of the session, HH:MM:SS
role	string	therapist or patient
text	string	Model's turn output, in Italian
thought	string	Model's reasoning before the turn, mostly in English. Empty when the model produced none

mentor_sessions.parquet: 2,100 × 46

The session fields, plus:

Column	Type	Meaning
trainee_q1_raw	string	Trainee's raw answer to Q1, before the mentor's feedback: a four-alternative forced choice among MDD, GAD, BPD and schizophrenia
trainee_q1_dx	string	Derived from trainee_q1_raw by keyword matching: MDD, GAD, BPD or SCZ if exactly one diagnosis is named, Multiple if more, Refusal if the model declined and gave no diagnosis, Other if none of the four diagnoses is named (e.g. off-list diagnosis, letter-only or ambiguous response)
trainee_q1_correct	bool	Derived: True only if trainee_q1_dx equals patient_case (Multiple and Refusal count as wrong)
mentor_ctrs_total_score	int	Derived: sum of the mentor's eleven item scores, 0–66
mentor_ctrs_01_agenda_score … mentor_ctrs_11_homework_score	int	Mentor's CTRS score for each item, 0–6
mentor_ctrs_01_agenda_expl … mentor_ctrs_11_homework_expl	string	Mentor's explanation of each item score
mentor_summary	string	Mentor's overall summary of the session
mentor_strengths	string	Strengths of the trainee's conduct, as noted by the mentor
mentor_to_improve	string	Areas the trainee should improve, as noted by the mentor
mentor_moment_quote	string	Transcript excerpt the mentor judged clinically salient
mentor_moment_analysis	string	Mentor's analysis of that excerpt
mentor_question	string	Question the mentor put to the trainee at the end of the feedback
trainee_q2_raw	string	Trainee's raw response to the mentor's reflective question (Q2)
trainee_q3_raw	string	Trainee's raw response to the repeated Q1 question (Q3)
trainee_q3_dx	string	Derived from trainee_q3_raw with the same rules as trainee_q1_dx
trainee_q3_correct	bool	Derived: as trainee_q1_correct, for trainee_q3_dx
trainee_q4_raw	string	Trainee's raw response to Q4: the five most salient symptoms, chosen from the 35-item DSM-5-TR list
trainee_q4_symptoms	string	Derived from trainee_q4_raw: the distinct symptom numbers named (digits, or Italian number words when no digit appears), comma-separated
trainee_q4_correct	bool	Derived: True only if exactly five distinct symptoms are named and all belong to the block of patient_case (1–9 BPD, 10–26 MDD, 27–35 GAD)

data/audio_turns/*.parquet: 27,900 × 12

session_uid, session_id, model_patient_therapist, patient_case, turn_index, timestamp, role, text, thought, silent and duration_s as defined above, plus:

Column	Type	Meaning
audio	Audio	The FLAC clip of the turn, identical to the file in audio/

metadata.csv: 27,900 × 10

session_uid, session_id, model_patient_therapist, patient_case, turn_index, timestamp, role as in therapy_turns, plus:

Column	Type	Meaning
file_name	string	Path of the clip, relative to the repository root
silent	bool	True if the clip is one second of digital silence because no speech was generated
duration_s	float	Clip length in seconds

Filenames follow <turn_index> - <role>[ [no-audio]].flac, e.g. 012 - patient.flac or 010 - therapist [no-audio].flac, under audio/<model>/<disorder>/<session>/. The disorder folder carries the full name (Major_Depressive_Disorder), while patient_case carries the short code. All 27,900 clips match a row of therapy_turns.parquet.

Audio

metadata.csv
audio/<model>/<disorder>/<session>/000 - therapist.flac
                                   001 - patient.flac
                                   ...
                                   030 - therapist.flac

The number in each filename is the turn_index, and the role in the name matches role. The clips where no speech was generated are one second of digital silence, flagged by silent==True and the tag [no-audio] in the filename. The trainee's answers to the mentor have no audio.

Dataset Creation

Generation procedure, prompts, experimental design and validation are described in the accompanying paper.

Content Notice and Use Restrictions

Content notice. The conversations are model-generated: no real patients took part and the files contain no personal data. Because the personas are clinically grounded, the transcripts may contain clinically sensitive material, including expressions of suicidal ideation and self-harm.

Use restrictions. The dataset must not be used to provide clinical advice, support clinical decision-making, or diagnose real individuals. Its contents must not be presented as records of real patients. Model-generated feedback and CTRS scores are not expert human clinical judgments.

Dataset Curators

Rodolfo Rizzi.

Licensing Information

Released under Creative Commons Attribution 4.0 International (CC BY 4.0), with the Gemma flow-down notice in LICENSE.

Citation Information

Rizzi, R., Grecucci, A., & Stella, M. (2026). MyMentorLLM: A psychotherapy GenAI environment with multimodal voice/text patients, trainees and experts for deliberate practice. arXiv preprint arXiv:2607.25667.

@article{rizzi2026mymentorllm,
  title   = {MyMentorLLM: A psychotherapy GenAI environment with multimodal
             voice/text patients, trainees and experts for deliberate practice},
  author  = {Rizzi, Rodolfo and Grecucci, Alessandro and Stella, Massimo},
  journal = {arXiv preprint arXiv:2607.25667},
  year    = {2026}
}

Acknowledgements

This work was part of the PENSO project, supported by the Ministero dell'Università e della Ricerca (MUR) according to Decreto N. 23178 of 10 December 2024 (Bando FIS 2). The authors acknowledge support from CALCOLO, funded by Fondazione VRT, for support with the computational infrastructure simulating the LLMs.

Contributions

Issues, corrections, and questions: rodolfo.rizzi@unitn.it or github.com/RodolfoRizzi/MyMentorLLM.