MentraSuite

Mengxi Xiao¹  Kailai Yang²  Pengde Zhao³  Enze Zhang¹  Ziyan Kuang⁴  Zhiwei Liu²  Weiguang Han¹  Min Peng¹  Qianqian Xie^1*  Sophia Ananiadou²

¹School of Artificial Intelligence, Wuhan University  ²The University of Manchester  ³School of Computer Science, Wuhan University  ⁴Jiangxi Normal University

Psychological Reasoning LLMs.

[ English | 简体中文 ]

Latest News

✨[2025.12.6] We released the hybrid model of our family, Mindora-chord. For downloading the model checkpoints, please click here: elsashaw/mindora-chord

✨[2025.9.4] We released the first model of our family, Mindora-r2. For downloading the model checkpoints, please click here: elsashaw/mindora-rl

Introduction

Mindora is a family of psychological reasoning LLMs designed for psychology-related tasks that demand strong reasoning abilities, including question answering, therapy plan generation, cognitive error analysis, and misinformation detection. We further evaluated the generalization ability of Mindora on unseen tasks such as psychiatric diagnosis and observed remarkable results.

Our base model is Qwen3-8B, and we obtained it through SFT and GRPO.

Benchmark

The script of MentraBench can be found in src/MentraBench.py. You can add your own model in the options of args.llm. Then you need to add your model in call_llm.py, and call your model in the get_llm() function of MentraBench.py.

You can evaluate your model by running the following script:

python MentraBench.py --llm [your_llm] --dataset_name [the_dataset_to_test]

Quick Start

You can use the model in the same way as using Qwen3-8B.

• Initialization

from modelscope import AutoModelForCausalLM, AutoTokenizer


class Mindora:
    def __init__(self, version):
        self.model = AutoModelForCausalLM.from_pretrained(
            model_name="elsashaw/mindora-rl2",
            torch_dtype="auto",
            device_map="cuda"
        )
        self.tokenizer = AutoTokenizer.from_pretrained(model_name)

    def generate(self, prompt):
        messages = [
            {"role": "system", "content": "Based on the following information of a case to make judgements. When answering, follow these steps concisely:\n\n 1. Reasoning Phase:\n   - Enclose all analysis within <think> tags\n   - Use structured subtitles (e.g., '###Comparing with Given Choices:') on separate lines\n   - Final section must be '###Final Conclusion:'\n\n2. Answer Phase:\n - Enclose your answer within <answer> tags\n - The answer phase should end with 'Answer: [option]'.\n - The answer should be aligned with reasoning phase. \nDeviation from this format is prohibited."},
            {"role": "user", "content": prompt}
        ]
        text = self.tokenizer.apply_chat_template(
            messages,
            tokenize=False,
            add_generation_prompt=True
        )
        model_inputs = self.tokenizer([text], return_tensors="pt").to(self.model.device)

        generated_ids = self.model.generate(
            **model_inputs,
            max_new_tokens=2048
        )
        generated_ids = [
            output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
        ]
        response = self.tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
        return response

• Usage example

mindora = Mindora()
response = mindora.generate(prompt="your prompt")

Acknowledgments

Model training is based on the LLaMA-Factory and VeRL frameworks.