metadata

license: cc-by-nc-nd-4.0
task_categories:
  - text-generation
  - text-classification
language:
  - ja
tags:
  - nlp
  - dialogue
  - dialogue-corpus
  - dialogue-system
size_categories:
  - 100M<n<1B

KokoroChat: A Japanese Psychological Counseling Dialogue Dataset Collected via Role-Playing by Trained Counselors

KokoroChat is the largest human-collected Japanese psychological counseling dialogue dataset to date (as of June 2025). It was created through role-playing between trained counselors and includes rich, long-form dialogues and detailed client feedback on counseling quality. The dataset supports research on empathetic response generation, dialogue evaluation, and mental health-oriented language modeling.

This work has been accepted to the main conference of ACL 2025. 📄 View Paper (arXiv)

🌟 Key Features

6,589 dialogues, collected between 2020 and 2024
Avg. 91.2 utterances per dialogue
480 trained counselors simulating online text-based counseling sessions
20-dimension Likert-scale client feedback for every session
Broad topic coverage: mental health, school, family, workplace, romantic issues, etc.

📊 Dataset Statistics

Category	Total	Counselor	Client
# Dialogues	6,589	-	-
# Speakers	480	424	463
# Utterances	600,939	306,495	294,444
Avg. utterances/dialogue	91.20	46.52	44.69
Avg. length/utterance	28.39	35.84	20.63

📁 Dataset Structure

Each sample contains:

A full counseling dialogue with role labels (counselor / client) and message timestamps
Structured client feedback on 20 dimensions (0–5 Likert scale)
Flags for ethical concern checks (optional)
Predicted topic label (automatically annotated by GPT-4o-mini)

👉 See the kokorochat_dialogues folder for the complete dataset.

🤖 Models Trained on KokoroChat

The following models were fine-tuned on subsets of the KokoroChat dataset with different feedback score ranges:

Llama-3.1-KokoroChat-Low:
Fine-tuned on 3,870 dialogues with client feedback scores < 70
Llama-3.1-KokoroChat-High:
Fine-tuned on 2,601 dialogues with scores between 70 and 98
Llama-3.1-KokoroChat-Full:
Fine-tuned on 6,471 dialogues (excluding 118 test samples with score ≥ 99), with scores ≤ 98

📄 Citation

If you use this dataset, please cite the following paper:

@inproceedings{qi2025kokorochat,
  title     = {KokoroChat: A Japanese Psychological Counseling Dialogue Dataset Collected via Role-Playing by Trained Counselors},
  author    = {Zhiyang Qi and Takumasa Kaneko and Keiko Takamizo and Mariko Ukiyo and Michimasa Inaba},
  booktitle = {Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics},
  year      = {2025},
  url       = {https://github.com/UEC-InabaLab/KokoroChat}
}

⚖️ License

KokoroChat is released under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0) license.