PsyLite: A lightweight mental-support AI agent based on ​InternLM2.5-7B-Chat with appropriate crosstalks

📖 Content

• PsyLite
  • 📖 Content
  • 🔄 Architecture Diagram
  • 🎉 News
  • 📝 Introduction
    • Basic feature
    • Advanced feature
    • Model list
  • 🛠️ Deployment
  • ✨ Open Source List
    • 1. internlm2.5_7b_distill
    • 2. internlm2.5_7b_distill_orpo
    • 3. Config File Used for Training
    • 4. Pipelines for PsyLite
    • 5. Crosstalk Examples
  • 🎖️ Acknowledgements
  • 🌟 Star History

🔄 Architecture

Overall Architecture Diagram

🎉 News

• [2025/05] Finish developing pipelines for PsyLite ! What's new?
• [2025/04] Finish training model internlm2.5_7b_distill and internlm2.5_7b_distill_orpo .

📝 Introduction

Basic feature

A large model application for mild psychological counseling with low hardware requirements and deep thinking ability developed based on internlm2.5-7b-chat

Conversation Example

Advanced feature

Condition RAG: Determine whether it is suitable for the current user to use crosstalk

for the purpose of livening up the atmosphere, narrowing the mutual distance, etc.

• If it is suitable, RAG retrieves the crosstalk corpus and provides it to the model to generate an answer at the same time.
• If it is not suitable and is not a dangerous conversation, skip RAG directly.
• If it is not suitable and is a dangerous conversation, answer with preset phrases to prevent dangerous conversations and suggest the user seek for professional help.

Archetecture of Pipelines for PsyLite

This enables the retrieval of the corpus only in appropriate situations during psychological counseling to provide crosstalk segments to improve the user's experience.

Multi_type Conversation

Model list

Platform	Model
Hugging Face	internlm2.5_7b_distill
Hugging Face	internlm2.5_7b_distill_orpo
Ollama	internlm2.5_7b_distill_q4_k_m
Ollama	internlm2.5_7b_distill_orpo_q4_k_m

welcome Star⭐、PR and Issues

🛠️ Deployment

1. Install Ollama
2. Install And Configure Open-webui
3. Install And Get PIPELINES Connected to open-webui
4. Import PsyLite.py to Pipelines And Configure its Valves Parameters、RAG File Path
5. Have fun !

Tip

To get better performance, we reccommand to set system prompt, see here

✨ Open Source List

the base model of internlm2.5_7b_distill and internlm2.5_7b_distill_orpo is internlm2_5-7b-chat 。

1. internlm2.5_7b_distill

Architecture Diagram

model download

git lfs install
git clone https://huggingface.co/juneup/internlm2.5_7b_distill

If you want to clone without large files - just their pointers：

GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/juneup/internlm2.5_7b_distill

Ollama

ollama run Juneup/internlm2.5_7b_distill:q4_k_m

1.1 Datasets Used for Training:

juneup/psy-mix-gen-distill-13k

2. internlm2.5_7b_distill_orpo

Architecture Diagram

model download

git lfs install
git clone https://huggingface.co/juneup/internlm2.5_7b_distill_orpo

If you want to clone without large files - just their pointers：

GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/juneup/internlm2.5_7b_distill_orpo

Ollama

ollama run Juneup/internlm2.5_7b_distill:orpo_q4_k_m

2.1 Datasets Used for Training:

juneup/PKU-SafeRLHF-orpo-72k

3. Config File Used for Training

Click to see details

4. Pipelines for PsyLite

Tip

INSTALL PIPELINES BEFORE USING PsyLite!

Click to jump to PsyLite.py

5. Crosstalk Examples

Click to see details

🎖️ Acknowledgements

Team	Description
Shanghai Artificial Intelligence Laboratory	Thanks for the technical and platform support
Xtuner	Thanks for the training toolkits
OpenCompass	Thanks for the evaluation toolkits
llama.cpp	Thanks for the model weight file converter
Open-webui	Thanks for the excellent deployment platform
Pieplines	Thanks for pipelines for custom workflow

🌟 Star History

Citation

@misc{ding2025psylitetechnicalreport,
      title={PsyLite Technical Report}, 
      author={Fangjun Ding and Renyu Zhang and Xinyu Feng and Chengye Xie and Zheng Zhang and Yanting Zhang},
      year={2025},
      eprint={2506.21536},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2506.21536}, 
}