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

💗 We introduce Kardia-R1: teaching LLMs to understand, reason, and care — with transparent empathy 🌱

💌 Contact: jamse_yuan@163.com

---

🔥 News

• 2026.03.03 🎉 Our Kardia-R1 (7B) model is now officially available on Hugging Face Model Hub! 🤗
• 2026.01.14 🎉 Excited to share that our work, Kardia-R1, has been accepted at WWW 2026!
• 2025.12.02 🎉 Our Kardia-R1 paper released on arXiv — check it out now!
• 2025.12.03 🚀 The full KardiaBench dataset (22K multi-turn dialogues, 671 personas) is now officially released and open-sourced on HuggingFace!

### Authentication & Loading the Dataset
huggingface-cli login
from datasets import load_dataset
dataset = load_dataset("Jhcircle/KardiaBench")

---

💞 What is Kardia-R1?

🧠 Kardia-R1 is a reasoning-centric empathetic dialogue framework that unifies
user understanding → emotional reasoning → safe, supportive responses,
empowered by Rubric-as-Judge GRPO Reinforcement Learning for transparent and controllable empathy.

---

🧩 Key Features

• KardiaBench: 671 real-world user profiles + 22,080 multi-turn empathetic dialogues
• Four-span structured cognition
  • <|understanding_begin|>...<|understanding_end|> — Interpret user intent & emotions using persona and context
  • <|reasoning_begin|>...<|reasoning_end|> — Perform internal appraisal and empathetic reasoning
  • <|emotion_begin|>...<|emotion_end|> — Identify the correct fine-grained emotion label
  • <|response_begin|>...<|response_end|> — Generate supportive, persona-aligned empathetic replies
• Rubric-as-Judge RL (Verifiable)
  • Interpretable, criterion-based, LLM-judged reinforcement learning
• Backbone-Agnostic Gains
  • Improves Qwen, Gemma, and more across all empathy metrics
• Superior to SoTA LLMs
  • Outperforms GPT-4o, DeepSeek-R1, PsyLLM in emotion accuracy & empathetic quality

---

🎯 Rubric-as-Judge RL (Verifiable Reinforcement Learning)

• Human-interpretable rubric: Relevance · Empathy · Persona Consistency · Safety · Fluency
• Transparent scoring → controllable improvement
• No black-box reward models → fully interpretable and aligned behavior

---

📈 Superior Performance

• Consistent gains across every empathy dimension
• Stronger emotional grounding and persona alignment
• Scalable to Qwen (3B/7B) / Gemma (2B/7B） backbones
• Robust, generalizable empathetic cognition across diverse emotional contexts

🌟 Kardia-R1 achieves state-of-the-art empathy, persona consistency, and emotion accuracy,
surpassing both general-purpose LLMs and specialized empathetic dialogue systems.

---

📦 KardiaBench Dataset

📂 Dataset Overview

KardiaBench is a large-scale empathetic dialogue dataset designed for
reasoning-centered emotional support, containing:

• 22,080 empathetic multi-turn dialogues
• 671 fully documented personas
• Fine-grained emotional states
• Four-span structured reasoning format
• Fully anonymized & cleaned

HuggingFace dataset: 👉 KardiaBench

---

📥 Load the Dataset

To prevent misuse of sensitive data, our dataset requires an access request on HuggingFace. Please follow the instructions below to obtain access.

• Submit an Access Request describing the intended use
• Wait for approval from the maintainers — we will review and approve requests as quickly as possible

from datasets import load_dataset

dataset = load_dataset("Jhcircle/KardiaBench")

If you encounter AccessDenied or 403 Forbidden errors, your access request may still be pending or your HuggingFace authentication may be missing.

Login manually if needed:

huggingface-cli login

📘Data Fields

Field	Description
person	Full raw user profile string including MBTI, About, Signature, and Recent Activities.
mbti	The user’s MBTI type extracted from the profile (e.g., “INFP”, “ISTP”).
emotion	Target emotional state representing the user’s current feelings in the scenario (e.g., “anxious”, “terrified”).
situation	Starting background context or emotional scenario for the conversation.
anon_username	An anonymized username for privacy-preserving user identity.
messages	Full structured dialogue as a list of message objects, including the system prompt, user turns, and assistant responses.

🚀 Quick Start (Kardia-R1)

Installation

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

Quick Start with Transformers (infer_hf.py)

import os
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
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 (infer_swift.py)

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}')

📚 Citation

If our work is helpful, please cite:

@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}
}

🙇 Acknowledgement

We gratefully acknowledge EmpatheticDialogues for foundational inspiration, PersonalityCafe for publicly shared personas, DeepSeek-R1 and Qwen3 for their GRPO insights, and all annotators and psychology experts for their invaluable support in building KardiaBench.