| --- |
| license: apache-2.0 |
| language: |
| - en |
| - zh |
| pipeline_tag: question-answering |
| tags: |
| - medical |
| --- |
| |
|
|
| # Leanly_AI |
| |
| ## Overview |
| |
| **Leanly_AI** is a family of domain-adapted large language models developed by researchers from the Department of Endocrinology and Metabolism and the Department of General Practice at Provincial Hospital Affiliated to Fuzhou University. |
| |
| The models are designed for psychological support and clinician–patient communication in obesity and clinical weight management settings. Leanly_AI aims to help physicians respond more consistently and empathetically to emotional difficulties that may arise during weight management, while maintaining medically cautious, structured, and clinically interpretable outputs. |
|
|
| Leanly_AI is the locally deployable text-model component of the broader **Leanly Agent** system. The text model generates supportive responses, practical behavioral suggestions, psychological risk reminders, and weight-management-related health information. The complete Leanly Agent workflow can additionally transform de-identified model outputs into physician reports, patient reports, health education materials, and illustrated educational tips. |
| |
| For additional information and deployment guidance, please visit: |
| |
| http://www.leanly-ai.top/ |
| |
| --- |
|  |
|  |
| ## Clinical Focus |
| |
| Weight management is often accompanied by emotional and behavioral challenges. Individuals with obesity may experience: |
| |
| * Anxiety about treatment outcomes or weight regain |
| * Low mood and frustration during weight-loss plateaus |
| * Guilt or self-blame after dietary lapses |
| * Emotional eating or perceived loss of control |
| * Body image distress |
| * Weight-related stigma |
| * Reduced confidence and self-efficacy |
| * Treatment fatigue and reduced motivation |
| * Sleep-related or social difficulties affecting weight management |
| |
| These problems do not always meet the diagnostic criteria for a mental disorder, but they may still affect eating behavior, physical activity, treatment adherence, follow-up attendance, and long-term engagement. |
| |
| Leanly_AI is designed to support clinicians in addressing these common emotional difficulties through non-judgmental communication, practical guidance, and appropriate risk reminders. |
|
|
| --- |
|
|
| ## From Generic Reassurance to Clinically Oriented Support |
|
|
| General conversational models may respond to emotional distress with broad encouragement such as: |
|
|
| > “Do not be sad.” |
| > “Keep going.” |
| > “You can do it.” |
|
|
| Although well intentioned, such responses may not adequately address the complex emotional burden experienced during obesity treatment. |
|
|
| Leanly_AI is designed to provide more structured support by: |
| |
| 1. Identifying the main emotional concern expressed by the user |
| 2. Acknowledging distress without judgment or blame |
| 3. Explaining possible relationships between emotions, eating behavior, sleep, physical activity, and treatment adherence |
| 4. Providing small and practical actions that can be implemented in daily life |
| 5. Avoiding stigmatizing, moralizing, or shame-inducing language |
| 6. Highlighting situations that may require further professional evaluation |
| 7. Reminding users that short-term weight fluctuations do not necessarily represent personal failure |
| |
| The model is intended to support clinical communication rather than provide generic emotional companionship. |
| |
| --- |
| |
| ## Model Development |
| |
| Leanly_AI was developed by performing supervised fine-tuning on Qwen3 base models. |
|
|
| ### Base models |
|
|
| The current model family includes variants based on: |
|
|
| * Qwen3-4B |
| * Qwen3-14B |
|
|
| Both Chinese and English models were developed. Thinking and non-thinking variants are available, resulting in eight model configurations across language, parameter size, and reasoning mode. |
|
|
| ### Training dataset |
|
|
| The supervised fine-tuning dataset contains approximately **2,100 question–answer pairs** focused on emotional support during obesity treatment and weight management. |
|
|
| The questions cover topics including: |
|
|
| * Anxiety and depressive emotions |
| * Emotional eating |
| * Weight-loss plateaus |
| * Fear of weight regain |
| * Body image concerns |
| * Social avoidance |
| * Sleep difficulties |
| * Self-blame and loss of confidence |
| * Reduced treatment motivation |
| * Difficulties maintaining dietary and physical activity plans |
|
|
| Approximately 53% of the questions were derived from de-identified questions raised by patients in clinical weight-management settings. The remaining questions were generated to broaden coverage of relevant emotional and behavioral scenarios. |
|
|
| The training answers were distilled from multiple teacher models and formatted according to a standardized clinical communication template. A subset of the generated answers was reviewed by mental health professionals. |
|
|
| ### Fine-tuning method |
|
|
| The models were trained using supervised fine-tuning with the LoRA parameter-efficient fine-tuning method through LLaMA-Factory. |
|
|
| The training objective was to improve the models’ ability to: |
|
|
| * Recognize emotional concerns relevant to weight management |
| * Provide supportive and non-stigmatizing responses |
| * Generate practical behavioral suggestions |
| * Summarize the main emotional issue |
| * Estimate the apparent severity of emotional distress based on the available text |
| * Provide appropriate reminders when professional psychological assessment may be needed |
|
|
| --- |
|
|
| ## Expected Output Structure |
|
|
| Depending on the model variant and prompt template, Leanly_AI can generate: |
| |
| * A supportive response addressing the user’s main concern |
| * A clinically oriented explanation of the emotional or behavioral problem |
| * Practical and achievable self-management suggestions |
| * Weight-management-related health education |
| * Six brief supportive tips |
| * A summary of the main emotional state |
| * An estimated level of emotional distress |
| * A short explanation supporting the estimated severity |
| * A recommendation for further assessment when high-risk information is explicitly present |
| |
| Risk-related conclusions must be based only on information contained in the user’s input. The model should not fabricate or infer suicidal ideation, self-harm, purging, uncontrolled binge eating, medication misuse, or other high-risk behaviors when they have not been explicitly described. |
| |
| --- |
| |
| ## Clinical Logic and Interpretability |
| |
| Leanly_AI emphasizes clinically interpretable communication rather than unrestricted conversational generation. |
|
|
| Its response strategy focuses on: |
|
|
| * Emotional difficulties that may interfere with weight management |
| * Psychological distress associated with weight stigma |
| * Body image concerns and internalized weight bias |
| * Shame, guilt, and self-blame following perceived treatment failure |
| * Emotional eating and treatment-related frustration |
| * Situations that may require psychological or psychiatric assessment |
| * Communication strategies that reduce judgment and stigma |
| * Small, practical actions that may support continued engagement in treatment |
|
|
| The model is intended to help physicians understand the structure of supportive communication. A typical response follows a clinically understandable sequence: |
|
|
| **Identify the concern → acknowledge the emotion → explain relevant mechanisms → provide practical actions → assess risk signals → recommend professional support when appropriate.** |
|
|
| --- |
|
|
| ## Weight-Stigma-Reducing Communication |
|
|
| A central design principle of Leanly_AI is the avoidance of weight-stigmatizing language. |
| |
| The model is trained to avoid expressions that portray obesity as evidence of: |
| |
| * Weak willpower |
| * Laziness |
| * Lack of discipline |
| * Moral failure |
| * Personal irresponsibility |
| |
| Instead, obesity is treated as a complex chronic condition influenced by biological, behavioral, psychological, social, and environmental factors. |
| |
| Leanly_AI encourages person-first, non-judgmental communication and aims to reduce shame-based interactions that may negatively affect trust, follow-up attendance, and treatment engagement. |
|
|
| --- |
|
|
| ## Local Deployment and Privacy |
|
|
| Leanly_AI is designed for local deployment. |
| |
| When deployed locally through tools such as Ollama and Open-WebUI: |
| |
| * Patient input can remain on the local computer |
| * The text model can operate without an internet connection |
| * Original patient information does not need to be transmitted to an external model provider |
| * Institutions can maintain greater control over model access and data handling |
| |
| The broader Leanly Agent system uses a hybrid architecture. Original patient input is first processed locally. Only content that has undergone privacy filtering and de-identification may be transferred to an online agent workflow for document generation or multimodal material production. |
| |
| De-identification reduces privacy risk but does not guarantee absolute anonymity. Institutional information-security policies, access controls, audit procedures, and human review remain necessary. |
| |
| --- |
| |
| ## Intended Uses |
| |
| Leanly_AI may be used as an auxiliary tool for: |
|
|
| * Supporting clinician–patient communication in weight-management clinics |
| * Drafting non-judgmental responses to emotional concerns |
| * Providing low-intensity, non-therapeutic emotional support |
| * Generating practical behavioral suggestions |
| * Supporting weight-stigma-reducing communication |
| * Identifying text-based signals that may require further clinical assessment |
| * Preparing educational or communication materials for physician review |
| * Training clinicians in structured and empathetic communication |
| * Conducting research on domain-specific medical language models |
|
|
| All clinically relevant outputs should be reviewed by a qualified healthcare professional before being used in patient care. |
|
|
| --- |
|
|
| ## Out-of-Scope Uses |
|
|
| Leanly_AI should not be used: |
| |
| * To independently diagnose depression, anxiety disorders, eating disorders, or other psychiatric conditions |
| * To replace psychiatrists, psychologists, physicians, or other qualified professionals |
| * To provide psychotherapy or psychiatric treatment |
| * To make autonomous referral or emergency decisions |
| * To determine whether a patient is safe without professional assessment |
| * To prescribe, stop, or adjust medication |
| * To generate final medical records without clinician review |
| * As the sole basis for managing suicidal ideation, self-harm, severe hopelessness, purging, uncontrolled binge eating, or medication misuse |
| * As a substitute for emergency services or established clinical pathways |
| |
| Individuals with acute self-harm or suicide risk, severe psychological distress, suspected eating disorders, or other urgent safety concerns require immediate assessment by qualified healthcare professionals. |
| |
| --- |
| |
| ## Relationship Between Leanly_AI and Leanly Agent |
|
|
| The two terms refer to different components: |
|
|
| ### Leanly_AI |
| |
| Leanly_AI is the locally deployable text-model family. It processes patient descriptions and generates supportive, structured, and clinically oriented text. |
|
|
| ### Leanly Agent |
|
|
| Leanly Agent is the complete clinical-support workflow built around Leanly_AI. After privacy filtering and de-identification, the system can use predefined agent skills to generate: |
| |
| * A structured physician report |
| * A patient-friendly support report |
| * Psychological risk reminders |
| * Follow-up and communication suggestions |
| * Six brief health education tips |
| * Illustrated educational materials |
| |
| Therefore, downloading Leanly_AI provides access to the text model, but not necessarily all document-generation and multimodal functions of the complete Leanly Agent workflow. |
|
|
| --- |
|
|
| ## Preliminary Evaluation |
|
|
| In the preliminary evaluation reported by the development team, domain fine-tuning improved the performance of the local Qwen3-4B and Qwen3-14B models on weight-management-related emotional-support tasks compared with their corresponding base models. |
|
|
| The thinking variants achieved higher median evaluation scores than the non-thinking variants and approached the performance of some larger online models on this specific task. |
|
|
| These findings should be interpreted cautiously because: |
|
|
| * The evaluation focused on a specific clinical communication task |
| * Part of the evaluation relied on an automated judge model |
| * The study was conducted by the development team |
| * Independent external validation has not yet been completed |
| * Text-quality performance does not establish clinical effectiveness |
| * The model has not been demonstrated to replace professional psychological assessment |
|
|
| The current evidence supports preliminary feasibility and task-specific usefulness rather than definitive clinical efficacy. |
|
|
| --- |
|
|
| ## Key Characteristics |
|
|
| ✅ Developed specifically for obesity and weight-management settings |
| ✅ Supports Chinese and English |
| ✅ Includes 4B and 14B model sizes |
| ✅ Includes thinking and non-thinking variants |
| ✅ Supports local and offline deployment |
| ✅ Uses non-judgmental and stigma-reducing communication |
| ✅ Generates structured and clinically interpretable responses |
| ✅ Provides practical, low-intensity emotional support suggestions |
| ✅ Includes reminders for potential psychological risk |
| ✅ Designed for physician-assisted use |
| ✅ Does not replace psychological or psychiatric professionals |
|
|
| --- |
|
|
| ## Important Safety Statement |
|
|
| Leanly_AI is a research and clinical communication support model. |
| |
| It is not a medical device, an autonomous diagnostic system, or a substitute for professional care. Its output may be incomplete, inaccurate, overly general, or unsuitable for a particular individual. |
| |
| All patient-facing materials, risk classifications, referral suggestions, and clinical reports generated with Leanly_AI or Leanly Agent must be reviewed and approved by a qualified healthcare professional. |
|
|
| --- |
|
|
| ## Core Summary |
|
|
| **Leanly_AI is a locally deployable, clinically oriented language-model family designed to support emotional communication during obesity treatment and weight management. It combines medical knowledge, real-world clinical scenarios, non-stigmatizing communication principles, and structured risk reminders to assist physicians in providing more consistent, empathetic, and practical support.** |
| |
| **Leanly_AI — Supporting more comprehensive, respectful, and sustainable weight-management care.** |
|
|
| --- |
|
|
|
|
| --- |
|
|
