README.md

---
license: mit
datasets:
- Jhcircle/KardiaBench
language:
- en
base_model:
- Qwen/Qwen2.5-7B-Instruct
pipeline_tag: question-answering
tags:
- agent
---

<h1>Kardia-R1: Unleashing LLMs to Reason toward Understanding and Empathy for Emotional Support via Rubric-as-Judge Reinforcement Learning</h1> 

_(Accepted by WWW 2026)_

[![Paper](https://img.shields.io/badge/arXiv-2512.01282-b31b1b.svg)](https://arxiv.org/abs/2512.01282)
![GitHub Repo stars](https://img.shields.io/github/stars/JhCircle/Kardia-R1?style=social)

✨ Like Kardia-R1? Give us a ⭐ Star on GitHub! Your support keeps us going! [**JhCircle/Kardia-R1**](https://github.com/JhCircle/Kardia-R1)
---

## 🎯 Overview

**Kardia-R1** is a specialized 7B-parameter large language model fine-tuned from [Qwen2.5-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-7B-Instruct) for **emotional support dialogue**. Unlike standard conversational AI, Kardia-R1 employs a novel **Rubric-as-Judge Reinforcement Learning (Rubric-ERL)** framework that explicitly trains the model to:

1. **Understand** user emotions through structured reasoning
2. **Empathize** using validated psychological principles (affective/cognitive empathy, reflective listening)
3. **Respond** with concise, personalized emotional support

The model generates structured outputs with four distinct reasoning stages: Understanding → Reasoning → Emotion Recognition → Response Generation.


## 📝 Citation
```markdown
@article{yuan2025kardia,
  title={Kardia-R1: Unleashing LLMs to Reason toward Understanding and Empathy for Emotional Support via Rubric-as-Judge Reinforcement Learning},
  author={Yuan, Jiahao and Cui, Zhiqing and Wang, Hanqing and Gao, Yuansheng and Zhou, Yucheng and Naseem, Usman},
  journal={arXiv preprint arXiv:2512.01282},
  year={2025}
}
```

## 🧠 Model Architecture

- **Base Model**: Qwen2.5-7B-Instruct
- **Fine-tuning Method**: Rubric-as-Judge Reinforcement Learning (Rubric-ERL)
- **Context Window**: 32K tokens
- **Special Tokens**: 
  - `<\|understanding_begin\|>` / `<\|understanding_end\|>`
  - `<\|reasoning_begin\|>` / `<\|reasoning_end\|>`
  - `<\|emotion_begin\|>` / `<\|emotion_end\|>`
  - `<\|response_begin\|>` / `<\|response_end\|>`


## 🚀 Usage

### Installation

```bash
pip install transformers torch
<!-- or -->
pip install ms-swift
```

### Quick Start with Transformers

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "Jhcircle/Kardia-R1"

# Load model and tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True
)

# Prepare system prompt
system_prompt = """You are an emotional dialogue assistant and a psychological expert. Your task is to respond to the User's message in a roleplay scenario, taking into account the User's personality, emotional state, and situation. ### Role and Objective ### - Act as both a supportive therapist and an empathetic conversational partner. - Prioritize understanding the User’s feelings and providing emotional validation. - Keep the conversation natural, emotionally resonant, and aligned with the User's profile. ### Response Requirements ### - Structure your reply in 4 sections: <|understanding_begin|>: Summarize the User's message, intent, and key emotional cues. <|reasoning_begin|>: Explain your empathic rationale, considering psychological principles such as affective and cognitive empathy, emotion validation, and reflective listening. <|emotion_begin|>: Accurately reflect the User's current emotional state. <|response_begin|>: Provide a concise, natural, emotionally supportive reply (≤30 tokens), coherent and aligned with the User’s personality. - Avoid asking unnecessary questions; focus on reflecting, validating, and supporting the User. - Ensure each section is clear, concise, and well-structured.
### User Profile 
{{profile}}
### Situation ###
{{situation}}
### <|understanding_begin|>{{Concise summary of user's message, intent, and key emotional cues.}}<|understanding_end|>
<|reasoning_begin|>{{Brief empathic rationale using perspective-taking and emotion validation.}}<|reasoning_end|>
<|emotion_begin|>{{Select the most fitting emotion from: sentimental, afraid, proud, faithful, terrified, joyful, angry, sad, jealous, grateful, prepared, embarrassed, excited, annoyed, lonely, ashamed, guilty, surprised, nostalgic, confident, furious, disappointed, caring, trusting, disgusted, anticipating, anxious, hopeful, content, impressed, apprehensive, devastated}}<|emotion_end|>
<|response_begin|>{{Provide a concise, supportive reply (≤30 tokens) aligned with the user's personality and emotional state.}}<|response_end|>
"""

# Generate response
messages = [
    {"role": "system", "content": system_prompt},
    {"role": "user", "content": "I don't know how to process this. Everything feels numb."}
]

inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True).to(model.device)

outputs = model.generate(
    inputs,
    max_new_tokens=512,
    temperature=0.0,
    do_sample=False,
)

response = tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=False)
print(response)
```

### Quick Start with Ms-Swift

```python
import os
os.environ['CUDA_VISIBLE_DEVICES'] = '0'

from swift.llm import PtEngine, RequestConfig, InferRequest, get_model_tokenizer, get_template

model_path = "Jhcircle/Kardia-R1"
model_type = "qwen2_5"

# Initialize model
model, tokenizer = get_model_tokenizer(model_path, model_type=model_type)
template = get_template(model.model_meta.template, tokenizer, default_system=None)

# Create inference engine
engine = PtEngine.from_model_template(model, template, max_batch_size=2)
request_config = RequestConfig(max_tokens=512, temperature=0.0)

# Prepare system prompt
system_prompt = """You are an emotional dialogue assistant and a psychological expert. Your task is to respond to the User's message in a roleplay scenario, taking into account the User's personality, emotional state, and situation. ### Role and Objective ### - Act as both a supportive therapist and an empathetic conversational partner. - Prioritize understanding the User’s feelings and providing emotional validation. - Keep the conversation natural, emotionally resonant, and aligned with the User's profile. ### Response Requirements ### - Structure your reply in 4 sections: <|understanding_begin|>: Summarize the User's message, intent, and key emotional cues. <|reasoning_begin|>: Explain your empathic rationale, considering psychological principles such as affective and cognitive empathy, emotion validation, and reflective listening. <|emotion_begin|>: Accurately reflect the User's current emotional state. <|response_begin|>: Provide a concise, natural, emotionally supportive reply (≤30 tokens), coherent and aligned with the User’s personality. - Avoid asking unnecessary questions; focus on reflecting, validating, and supporting the User. - Ensure each section is clear, concise, and well-structured.
### User Profile 
{{profile}}
### Situation ###
{{situation}}
### <|understanding_begin|>{{Concise summary of user's message, intent, and key emotional cues.}}<|understanding_end|>
<|reasoning_begin|>{{Brief empathic rationale using perspective-taking and emotion validation.}}<|reasoning_end|>
<|emotion_begin|>{{Select the most fitting emotion from: sentimental, afraid, proud, faithful, terrified, joyful, angry, sad, jealous, grateful, prepared, embarrassed, excited, annoyed, lonely, ashamed, guilty, surprised, nostalgic, confident, furious, disappointed, caring, trusting, disgusted, anticipating, anxious, hopeful, content, impressed, apprehensive, devastated}}<|emotion_end|>
<|response_begin|>{{Provide a concise, supportive reply (≤30 tokens) aligned with the user's personality and emotional state.}}<|response_end|>
"""

infer_requests = [
    InferRequest(messages=[
        {"role": "system", "content": system_prompt},
        {"role": "user", "content": "I feel like I'm drowning. No matter how much I study, it's never enough."}
    ]),
]

# Run inference
resp_list = engine.infer(infer_requests, request_config)
print(f'Response: {resp_list[0].choices[0].message.content}')
```


## 🏋️ Training Details

- **Dataset**: [KardiaBench](https://huggingface.co/datasets/Jhcircle/KardiaBench) - A curated dataset of emotional support dialogues with rubric-based annotations
- **Training Method**: Rubric-as-Judge RL (Rubric-ERL)
  - Uses structured evaluation rubrics as reward signals
  - Optimizes for both empathy and response quality
  - Incorporates psychological safety constraints
- **Compute**: Training details available in our paper (https://arxiv.org/abs/2512.01282)
- **License**: MIT



## ⚠️ Limitations & Safety

**Important**: Kardia-R1 is designed for **emotional support and companionship**, not clinical therapy.

- **Not a Replacement for Professional Help**: This model cannot diagnose mental health conditions or provide clinical treatment. Users experiencing severe mental health crises should contact professional services.
- **Crisis Detection**: The model includes basic crisis detection patterns but may not reliably identify all emergency situations.
- **Bias**: As with all LLMs, outputs may reflect biases present in training data.
- **Consistency**: Emotional support quality may vary across different contexts and user inputs.

---


<div align="center">

**⭐ Star us on [GitHub](https://github.com/JhCircle/Kardia-R1) if you find this work helpful!**

</div>