dzur658/therapy-conversations-full-small

Therapy Conversations Full Small

About the dataset

This dataset contains 3161 unique, synthetically generated examples, of multi-turn conversations between a patient and a therapist. Minimax M 2.5 was used to generate the transcripts, but there is also associated meta data attached. A second and third pass by MiniMax M 2.5 over the original data was used to extract both entities, and extract relationships between those entities.

Understanding an Object in the Dataset

Top Level Keys

• —conversation contains the particular conversation in json format
• —fingerprint contains metadata about the conversation (demographic, personality traits of both parties, etc)
• —full text contains the full text of the transcript in the way it woul be ingested into the Knowledge graph genrator API
• —knowledge_graph_shard contains the entities and relations MiniMax M 2.5 extracted

conversation

conversation contains a turns object. The turns object contains sub-objects that are each "turn" of the conversation. Each turn is made up of 2 keys speaker and text. Use this if your trying to inspect lines individually within a conversation, it is pre-parsed for you

fingerprint

fingerprint contains the following metadata about the conversation:

• —age: 18-75
• —gender: male, female, transgender woman, transgender man, non-binary
• —occupation: software engineer, teacher, nurse, artist, salesperson, retired, student, unemployed, entrepreneur, stay-at-home parent
• —presenting_issue: anxiety, depression, relationship issues, work stress, grief, self-esteem issues, sexuality issues, trauma, substance abuse, eating disorders, chronic illness, gender dysphoria, identity issues, family conflict, life transitions, body dysmorphia, obsessive-compulsive disorder, phobias, sleep disorders, anger management issues
• —relationship_status: single, in a relationship, married, divorced, widowed
• —living_situation: living alone, living with family, living with roommates, living with partner, living in a group home, living in a shelter
• —therapy_modality: cognitive-behavioral therapy, psychodynamic therapy, humanistic therapy, integrative therapy, mindfulness-based therapy, art therapy, dialectical behavior therapy, acceptance and commitment therapy, eye movement desensitization and reprocessing (EMDR), exposure therapy
• —patient_speaking_style: verbose, concise, emotional, logical, intellectualizing, narrative, disorganized, rambling, focused, tangential, metaphorical, literal
• —therapist_speaking_style: empathetic, direct, analytical, supportive, challenging, reflective, encouraging, neutral, collaborative, authoritative, offensive, dismissive, condescending, patronizing, invalidating
• —session: 1-20 (how many sessions the patient and therapists have been seeing each other)
• —conversation_length: the requested conversation length (although this most likely does not match as it was not strictly enforced during generation)

knowledge_graph_shard

knowledge_graph_shard contains two subkeys: entities and relations. Entities were identified by MiniMax M 2.5 on a second pass (first pass refers to transcript generation itself), and relationships were created from those entities in the relations key on a third pass from MiniMax M 2.5.

entities

• —text: the exact text from the transcript MiniMax M 2.5 identified to be an entity on it's second pass
• —label: Can be Symptom, Trigger, Emotion, Person, Coping_Mechanism, Life_Event, or Behavior
• —start: character index relating to the transcript where the start of the extracted entity can be found
• —end: character index relating to the transcript where the end of the extracted entity can be found
• —id: follows the schema Ex where x is the count for entities extracted from the transcript

relations

• —source: plaintext of the source entity
• —source_id: entity number (following the Ex schema)
• —predicate: the relationship between the source and target entity. Can be: CAUSES, WORSENS, IMPROVES, RELATES_TO, EXPERIENCES, or TRIGGERS
• —target: plaintext of the target entity
• —target_id: entity number (following the Ex schema)
• —proposed_by: Identifies which party first proposed the relationship in the transcript. Can be either Patient or Therapist
• —patient_acceptance: Identifies how the patient reacted to the information. If the patient proposed the relationship then this field will always be Affirmed. If the therapist proposed the relationship the value can be Affirmed, Denied, Avoided, or Realized_Later

Synthetic Data Generation Script

If you'd like to learn more about how this data was generated you can view the generation code here

Models Used

The following models were used to generate this data:

• —MiniMax M 2.5 Q_8

Citation

If you find this dataset useful, please cite me. :)

@misc{therapy-conversations-full-small,
    title  = {{therapy-conversations-full-small}},
    author = {{Alex Dzurec}},
    month  = {March},
    year   = {2026},
    url    = {https://huggingface.co/dzur658}
}