🧠 Llama-3.1-KokoroChat-Low: Japanese Counseling Dialogue Model

Llama-3.1-KokoroChat-Low is a large-scale Japanese language model fine-tuned on the entire KokoroChat dataset—a collection of over 6,000 psychological counseling dialogues conducted via role-play between trained counselors. The model is capable of generating empathetic and context-aware responses suitable for mental health-related conversational tasks.

---

💡 Overview

• ✅ Fine-tuned on 3,870 dialogues with client feedback scores below 70
• ✅ Data collected through text-based role-play by trained counselors
• ✅ Covers a wide range of topics: depression, family, school, career, relationships, and more
• ✅ Base Model: tokyotech-llm/Llama-3.1-Swallow-8B-Instruct-v0.3

---

⚙️ Usage Example

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "UEC-InabaLab/Llama-3.1-KokoroChat-Low"

# Load tokenizer and model
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")

# Set pad_token_id
if tokenizer.pad_token_id is None:
    tokenizer.pad_token = "[PAD]"
    tokenizer.pad_token_id = tokenizer.convert_tokens_to_ids("[PAD]")

model.config.pad_token_id = tokenizer.pad_token_id

# Build dialogue input
messages = [
    {"role": "system", "content": "心理カウンセリングの会話において、対話履歴を考慮し、カウンセラーとして適切に応答してください。"},
    {"role": "user", "content": "最近、気分が落ち込んでやる気が出ません。"}
]

# Tokenize with chat template
inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt"
).to(model.device)

attention_mask = inputs.ne(tokenizer.pad_token_id)

# Generate response
outputs = model.generate(
    inputs,
    attention_mask=attention_mask,
    pad_token_id=tokenizer.pad_token_id,
    max_new_tokens=256
)

# Extract only the newly generated tokens
response = outputs[0][inputs.shape[-1]:]
response_text = tokenizer.decode(response, skip_special_tokens=True)

# Print clean response
print(response_text)

---

🛠️ Fine-Tuning Details

Fine-tuning was performed using QLoRA with the following configuration:

• Quantization: 4-bit NF4 with bfloat16 computation
• LoRA target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• LoRA parameters:
  • r = 8
  • lora_alpha = 16
  • lora_dropout = 0.05

Dataset Split

• Training Data: 3,870 dialogues with feedback scores < 70
• Train/Validation Split: 90% train, 10% validation

Hyperparameter Settings

• Optimizer: adamw_8bit
• Warm-up Steps: 100
• Learning Rate: 1e-3
• Epochs: 5
• Batch Size: 8
• Validation Frequency: every 400 steps

---

📄 Citation

If you use this model or 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}
}

---

🔗 Related

• 📁 Dataset:
  • KokoroChat on Hugging Face Datasets
  • KokoroChat on GitHub (UEC-InabaLab)
• 🤖 Model Variants:
  • Llama-3.1-KokoroChat-High: fine-tuned on 2,601 dialogues with client feedback scores between 70 and 98
  • Llama-3.1-KokoroChat-Full: fine-tuned on 6,471 dialogues with client feedback scores ≤ 98
• 📄 Paper: ACL 2025 Paper (arXiv)