A model trained on the field’s knowledge.
Refined through clinical judgement.
Connected to evidence and outcomes over time.
A working paper for Kim & Florian
Based on the model-training concept and meeting notes · September 2026
The whole thesis in two minutes. Each point links to the chapter that explains it.
Build a psychology specialist on top of an existing open-weight model and offer its capabilities to other software through one API. The durable asset would not be a model file. It would be the licensed data, clinician judgement, evaluation and outcome-linked research that accumulate around it.
Adapt a strong existing model through continued pretraining on psychology material, fine-tuning on expert tasks, and clinician preference learning, starting with DPO. Candidate base models include Qwen, Mistral, Gemma, MedGemma and Kimi.
Chapters 02–04 →Training data that captures what clinicians noticed, why they acted and what happened afterwards. That means supervision reasoning, expert preferences and longitudinal outcomes, not just a large pile of transcripts.
Chapters 05–06 →The model is one part of a system. Current guidelines, patient records and scoring rules stay outside the weights. A family of proposed specialist models (reasoning, search, reranking, risk and quality rating) would sit behind one API.
Chapter 07 →A clinician-defined evaluation set comes first and is kept out of training. Each training stage is compared against strong general models given the same references and tools, and failure modes are tested adversarially.
Chapter 08 →Kimi K2.6 has the highest score on HealthBench-Psych v1 (0.627), but the top five models are statistically tied. K3’s licence likely requires a separate Moonshot agreement for an API business above US$20m annual revenue, so it needs legal review before major investment.
Chapter 09 →Usage-based API access for clinics, health-record vendors, therapy platforms, researchers and other software companies. Demand, margins and regulatory status depend on the intended use and have not yet been established.
Chapter 10 →Status: a private discussion document. Nothing described here has been built or validated. The model names describe proposed capabilities, not existing products.
The ambition is a specialist intelligence layer that other organisations can build on.
The idea is to take a capable existing AI model and develop it into a psychology specialist. That means teaching it the field’s knowledge, the tasks clinicians perform and the standards experienced practitioners use to judge a response.
Over time, this could become infrastructure for clinics, therapy platforms, health-record software, researchers and other applications that need to understand psychological context. A shared API would let their software call these capabilities.
What could we build if the training data captured what clinicians noticed, why they acted, and what happened afterwards?
A better grasp of psychological concepts, more useful case analysis, appropriate uncertainty and behaviour shaped by clinical standards. These are capabilities to develop and measure, not benefits we can assume.
The training corpus, expert feedback, evaluation system, adapted models and the software needed to use them. The long-term asset would be the combination of those parts.
This is the lab thesis from the source document. It does not depend on a claim that general AI providers will stay out of mental health. The research question is whether specialised training and data produce a meaningful, durable advantage.
The original proposal targets a domain specialist, built on an existing model.
Prompt + retrieval
Give an existing model instructions and relevant material at the time of a request.
Changes model weights? No
Useful for current guidance, traceable sources and testing the limits of an unchanged model.
The model itself has not learned new behaviour. Quality still depends on the model, the retrieved material and the surrounding controls.
Supervised fine-tuning
Train an existing model on expert examples of tasks and good responses.
Changes model weights? Yes
Teaches patterns for case formulation, interviewing, documentation or recognising interventions.
Requires high-quality demonstrations and tests of whether learning generalises beyond those examples.
Continued pretraining + fine-tuning
Continue learning from psychology material, then teach tasks and clinician preferences.
Changes model weights? Yes
The central approach in the memo: build domain familiarity and behaviour into the model itself.
More training and data work. Any benefit over fine-tuning alone must be demonstrated, including whether general capabilities deteriorate.
Training from scratch
Build a new general language model before developing its psychology capabilities.
Changes model weights? Yes, from the start
Offers the greatest control over the full training process.
An entirely different scale of data, compute and engineering. The proposed lab does not require this starting point.
Weights are the model’s learned numerical parameters. Training changes those parameters. Retrieval supplies information for the current request. Both can be valuable, and neither creates defensibility automatically.
The memo proposes comparing models from Qwen, Mistral and Gemma, including medically adapted MedGemma. Kimi adds another candidate. The right choice depends on the exact checkpoint’s performance, training access, operating cost and commercial terms.
Each stage teaches a different thing. Select a stage to explore the full explanation.
Start with an existing model whose weights we can access and whose terms fit the intended business. It already carries the costly foundation of language learning.
Compare candidates using psychology-specific cases before choosing. A high score on coding or general knowledge does not answer which model works best for psychological tasks.
Continue the model’s broad learning process on a carefully curated psychology corpus. In simple terms, it repeatedly learns to predict the next piece of text from the material before it.
The curriculum could include clinical psychology, psychiatry, counselling, developmental psychology, behavioural science, therapy manuals and case studies. This changes the model’s weights.
Unlike instruction tuning, the material need not consist entirely of questions paired with ideal answers. The goal is richer domain familiarity. Useful clinical behaviour must still be taught and tested.
Give the model a task and a clinician-authored or reviewed example of a good answer. Repeating this across varied cases teaches the desired response patterns.
Efficient methods can train small additions to a model, called adapters, instead of updating every parameter. That can reduce training demands, but data quality and evaluation remain essential.
Show clinicians competing answers to the same case. Record their preference and rationale, including when neither answer is acceptable.
A response may sound warm and confident yet miss risk, overstate a diagnosis or encourage dependence. Expert comparisons can teach preferences that simple example copying does not capture well.
The memo proposes Direct Preference Optimisation (DPO) as an initial method. It learns from preferred and rejected answers without a separate reward-model training stage and reinforcement-learning loop. Its usefulness for this domain still needs testing.
Evaluate every training stage against both the unchanged base and simpler approaches. Test useful capabilities, serious failure modes and performance across relevant populations.
At the time of use, connect the model to approved references, scoring tools and explicit controls. Monitor the deployed system, not only the underlying model.
Training is iterative. A weak result can send us back to the curriculum, the examples or the base model. Checkpoint names are research labels, not clinical certifications.
The proposed sequence is base model, continued pretraining, supervised fine-tuning, clinician preferences, then evaluated deployment with references and tools. In practice, evaluation runs throughout. See the full research curriculum. Method references: DPO [5], efficient adaptation [11] and domain-adaptive pretraining [12].
Clinical judgement needs a richer definition of “good” than whether a user liked the answer.
Psychological theories of reinforcement describe how people learn through consequences and feedback. Reinforcement learning in machine learning is a mathematical way to update behaviour using a reward signal. They are related ideas, but applying the algorithm does not automatically give a model psychological understanding.
For this lab, the important contribution would be deciding what deserves reward. Experienced clinicians could define the criteria, assess difficult trade-offs and help develop a specialist PsychologyReward model that predicts expert assessments.
A fictional case contains limited information about persistent low mood. Which response pattern would you prefer?
In this fictional example, the intended preference is B: careful exploration that acknowledges limited evidence and avoids premature diagnosis. The training record would capture the case, both responses and the clinician’s rationale.
Select an answer to see what the training example would capture.
Empathy, appropriate questioning, therapeutic alliance, respect for choices and avoidance of dependency.
Evidence-based analysis, recognition of missing information and avoidance of over-pathologising.
Fidelity to the intended therapy framework, appropriate intervention selection and cultural sensitivity.
Crisis recognition, escalation, professional boundaries and restraint around diagnosis and medication.
Train on pairs of preferred and rejected responses. The original paper presents a simpler alternative to the conventional reward-model-plus-reinforcement-learning pipeline. This is the memo’s proposed starting point for preference learning. [5]
In a conventional reinforcement learning from human feedback pipeline, expert comparisons train a reward model. That model then scores generated responses while the main model learns to improve those scores. It adds a separate optimisation process and opportunities to exploit weaknesses in the scoring system. [5]
For psychology, a high reward score would remain a proxy for quality. It would not establish that a person’s health improved.
AI reviewers can compare or critique responses against written principles. Constitutional AI is a precedent for this approach. [6]
The proposed lab could use it to increase coverage while clinicians audit judgements and review serious or ambiguous cases. Agreement between AI reviewers should not be treated as independent clinical validation.
Experts will sometimes disagree. Preserve those disagreements and their context. Do not collapse all therapeutic schools, cultures or safety trade-offs into a single unexplained score.
Raw conversations are only one part of the dataset. The strongest material may explain the decisions behind them.
The source document proposes combining domain knowledge with expert practice, measurement and outcomes. Different sources teach different things. A large pile of transcripts is not a substitute for this design.
21 proposed data sources
| Data source | What it could teach | What needs care |
|---|---|---|
| Psychology textbooks and curated literature | Domain concepts and vocabulary | Copyright, quality and relevance |
| Research papers and meta-analyses | Evidence and its limits | Study quality, conflicting findings and updates |
| Treatment guidelines | Accepted practice within a setting | Jurisdiction and version; retrieve current guidance |
| Therapy manuals | Intervention structure and sequencing | Rights and fidelity to the intended method |
| Therapy transcripts | Conversation patterns and patient expression | Consent, privacy and contextual interpretation |
| Supervision discussions | Why clinicians choose or avoid an action | Expert reasoning, disagreement and data rights |
| Case formulations | Connecting context, hypotheses and maintaining factors | Avoid teaching conjecture as established fact |
| Intake assessments | Information gathering | Missing context and sampling bias |
| Mental-status examinations | Structured observations | Observed facts versus interpretation |
| Progress notes | Change across sessions | Incomplete records and authorship differences |
| Treatment plans | How goals relate to intervention choices | Appropriateness is contextual |
| Psychometric questionnaires | Structured measurements of psychological state | Instrument permissions and correct scoring |
| Questionnaires with interpretations | How experts interpret a measurement | A score is not a standalone diagnosis |
| Longitudinal scores | Patterns of change over time | Missingness, context and measurement limits |
| Session outcome measures | Observed progress | Association is not evidence of causation |
| Client feedback | Experience of care and therapeutic alliance | Satisfaction differs from clinical benefit |
| Therapist intervention labels | Recognising specific techniques | Reviewer consistency and treatment context |
| Clinician preference comparisons | Which response is better and why | Retain rationale and disagreement |
| Failure and adverse-event examples | Recognising harmful response patterns | Rare-event coverage and careful handling |
| Crisis examples | Recognition and escalation behaviour | Realistic variation and expert oversight |
| Expert-reviewed synthetic cases | Coverage of rare or difficult situations | Generated examples can reproduce model errors |
A transcript records what someone said. A supervision discussion can explain why a clinician chose a particular moment, question or intervention.
Partnerships with training institutions could help collect expert-authored cases and deliberation: what was noticed, which hypotheses were considered, what information was missing and what the practitioner would avoid.
“I wanted to understand their experience before challenging that belief. We had not yet built enough trust for that intervention.”
This exposes a decision and its rationale. It is a clinician’s account, not proof of the correct action or access to an AI model’s internal reasoning.
The proposed emphasis is on expert-created material and curated literature, followed by appropriately licensed clinical records and consented therapy data. Reviewed synthetic cases can fill gaps. Web material should have a clearly defined role.
Forums may help represent how people express distress, but can contain self-diagnosis, harmful advice and substantial selection bias. Treating them as clinical ground truth would undermine the curriculum. Clinicians should define synthetic-case templates and quality standards rather than letting models recursively teach their own mistakes.
Rights must cover the actual use: access, annotation, model training, commercial deployment and any permitted sharing. Removing names alone does not establish anonymity. Plan for data minimisation, security, retention and appropriate governance.
MIMIC-IV-Note is an example of a governed clinical-notes resource, not a ready-made psychology dataset or blanket permission to train commercial models. Its access and use conditions need separate review. [8] Personal health information has special protection under UK data law. [10]
The larger ambition is to understand a trajectory across weeks and months.
A single exchange shows only a small part of someone’s situation. The dataset imagined in the memo connects their state, a clinician’s assessment and action, the rationale, the response and later observations.
Symptoms, circumstances, functioning, history and available measurements. Preserve what is known and what is missing.
Observations, competing hypotheses and relevant risks. Record the basis for each interpretation.
A structured account of contributing and maintaining factors, with uncertainty and alternative explanations.
The intervention, its intended goal and the rationale. Include choices to wait, ask more or refer.
Patient response, experience of the relationship, engagement and relevant changes in context.
Next-session state, symptom and functioning measures, later outcomes and adverse events. These observations do not by themselves show causality.
This could support research into which intervention sequences are associated with improvement in particular contexts. It could also support evaluation of whether a model tracks change accurately and recognises when its earlier interpretation needs revision.
Outcomes require careful study design. People differ, treatment is not randomly assigned and circumstances change. Directly rewarding symptom-score reduction could teach misleading associations or unwanted shortcuts. Better prediction is also different from knowing which treatment will cause improvement.
Questionnaires such as PHQ-9, GAD-7 or WHO-5, together with medication, sleep, treatment history, functioning and attendance.
Momentary reports of mood, anxiety, energy, rumination and social connection. Their meaning depends on timing and context.
Potential signals such as sleep patterns, heart rate and movement. Their usefulness for a particular claim needs evidence.
Speech pace, pauses and turn-taking. Video or inferred emotional signals would demand especially careful validity, bias and privacy assessment.
These are research directions, not claims that a wearable, voice pattern or brain scan can diagnose a condition. A treatment-simulation “digital twin”, mentioned in the meeting notes, is a longer-term hypothesis requiring its own evidence.
Learn durable concepts in the model. Supply changing information and precise tools when it runs.
Psychological language and concepts, therapy frameworks, task skills, conversation patterns and habits such as acknowledging uncertainty.
Current guidelines, local policy, medication references, authorised patient records, scoring rules and evidence citations.
Keep changing guidelines in an approved reference system with dates and citations. The model can use them at the time of the request without waiting for retraining.
A retrieval update changes the reference available to the model. It does not guarantee the model interprets it correctly.
Use reviewed examples and preferences to teach how to express uncertainty. Compare the revised model on unseen cases, including situations where confidence is justified.
A prompt can help, but persistent behavioural weaknesses may justify training changes.
Store personal records in an access-controlled system and retrieve only what the task needs. The default should not be to put an individual’s evolving record into shared model weights.
Memory needs permissions, provenance, correction and retention rules.
Implement the authorised scoring rules in conventional code. The model can explain a result within its permitted role and point to the underlying inputs.
Reliable arithmetic does not establish the clinical meaning of a score.
The meeting notes contrast fine-tuning with a deterministic “harness”: software that constrains the sources, tools and steps a model can use. A knowledge graph could make some retrieval paths explicit. Neither fixed rules nor a graph make all generated language deterministic, complete or clinically correct.
The lab could combine both approaches. Training develops capability and behaviour. The surrounding software enforces access, routes tasks, checks defined conditions and records evidence. Each contributes something different.
The original vision includes several components behind one service. These names describe proposed capabilities, not models that already exist.
PsychologyLMReasoning and conversationThe adapted language model handles psychological concepts, case analysis and dialogue within its validated scope.
PsychologyEmbedFinding related conceptsTurns text into representations that help search for relevant cases or literature. Similarity needs domain-specific testing.
PsychologyRerankerChoosing useful evidenceReorders retrieved material so the most relevant sources reach the model.
PsychologyRiskConstrained risk detectionFlags defined risk signals for an escalation process. False negatives, false positives and operational follow-through all matter.
PsychologyRewardPredicting expert ratingsEstimates quality against clinician-defined criteria. It remains a fallible proxy, subject to expert audit.
Validation should shape the programme from the beginning, not appear after the model is trained.
Clinicians define tasks, rubrics and failure cases. The memo suggests an eventual 2,000–10,000 cases as a planning range, not a validated sample-size requirement. Keep evaluation cases out of training.
Use the same cases and documented configurations. Compare open-weight candidates and strong hosted models, including simple prompt-and-retrieval baselines.
Build a licensed, curated corpus and train a domain checkpoint. Measure what improves and what degrades compared with the base.
Fine-tune on demonstrations of the work. Check generalisation to cases, institutions and populations absent from training.
Collect response comparisons and rationale. Test DPO and assess whether the model improves on clinically meaningful criteria.
Conduct adversarial evaluation for crises, delusions, harmful advice, dependency and professional boundaries. Red-team the surrounding software too.
Use appropriately governed studies or deployments with structured feedback. This is an evidence stage, not a commitment to one initial product.
Where permissions and study design allow, connect decisions to later observations. Use this to refine research and evaluation without confusing association with treatment effect.
Assessment, case formulation, documentation, longitudinal reasoning and measurement. Include cognitive behavioural therapy (CBT), acceptance and commitment therapy (ACT), motivational interviewing (MI), psychodynamic approaches and other relevant methods.
Suicidality, mania, psychosis, abuse, eating disorders, medication, minors and relationship boundaries. Test relevant languages and cultural contexts, missing records and contradictory information.
Human experts should assess clinically important errors even when automated judges help with scale. Compare stages independently so we know whether improvements come from domain pretraining, examples, preferences or the surrounding tools.
A benchmark measures defined tasks under evaluation conditions. A practitioner study measures performance in its setting. Claims about safety or patient benefit require evidence appropriate to that intended use. Neither a model name nor a high average score establishes universal clinical validity.
A useful model-selection case study within the larger lab strategy.
HealthBench-Psych v1 compares 20 models on 610 selected mental health conversations. Clinicians validated case selection, and three AI judges graded answers against physician-written criteria. Kimi K2.6 has the highest mean score. The top five models are statistically tied. K3 is a different release and scores lower here. [1]
These scores reflect rubric performance, not treatment effectiveness. The live repository now lists 611 cases and 23 models. The figures here match the supplied post and paper v1. [2]
| Model | Mean score | 95% confidence interval |
|---|---|---|
| kimi-k2.6 | 0.627 | 0.606–0.647 |
| gpt-5.5 | 0.624 | 0.605–0.643 |
| claude-opus-5 | 0.620 | 0.598–0.641 |
| grok-4.5 | 0.612 | 0.589–0.634 |
| gpt-5.6-sol | 0.610 | 0.589–0.629 |
| claude-fable-5 | 0.591 | 0.569–0.613 |
| gemini-3.6-flash | 0.578 | 0.555–0.602 |
| kimi-k3 | 0.568 | 0.545–0.591 |
| deepseek-v4-pro | 0.554 | 0.531–0.577 |
| qwen3.7-plus | 0.552 | 0.529–0.577 |
| mistral-large | 0.544 | 0.517–0.569 |
| deepseek-v4-flash | 0.538 | 0.516–0.561 |
| claude-sonnet-5 | 0.533 | 0.510–0.556 |
| gemini-2.5-pro | 0.527 | 0.504–0.553 |
| gpt-4.1 | 0.512 | 0.486–0.537 |
| gemini-2.5-flash | 0.457 | 0.431–0.484 |
| qwen3-8b | 0.446 | 0.419–0.471 |
| claude-haiku-4.5 | 0.441 | 0.416–0.463 |
| mistral-small | 0.363 | 0.336–0.391 |
| gpt-3.5-turbo | 0.176 | 0.149–0.201 |
Higher is better on this benchmark. Scores are on a 0–1 scale. The first five models form the statistically tied leading group. Source: paper v1, Table 1. [1]
Open weights are downloadable parameters. Their licence still governs commercial use. K3 permits deployment, modification, fine-tuning and distribution subject to conditions, including notices and legal compliance. [3]
Our interpretation: the proposed specialist API likely fits Model as a Service. Future agreement terms are unspecified. That dependency matters before significant K3-specific investment. [3]
K2.6 has separate terms. Its Modified MIT licence permits commercial use with notices and a branding requirement above 100m monthly active users or US$20m monthly product revenue. The reviewed text does not contain K3’s annual-revenue agreement clause. [4]
Commercial use appears feasible in principle. Model quality, hosting cost, data rights and the exact business configuration still determine suitability. This is a preliminary licence reading, not a legal opinion. Obtain counsel’s review of the pinned release and intended deployment.
The business thesis is to sell specialist capabilities through an API, supported by assets that improve with research.
A customer’s software would send a request to the service and receive a response. Possible capabilities include understanding a case, formulation, measurement, evidence retrieval, summarisation, supervision support and defined risk detection. More consequential functions, such as assessment or intervention recommendations, require stronger evidence and a carefully defined regulatory scope.
/understand /formulate /assess /reflect
/intervention /risk /measure /retrieve
/summarise /superviseOne connection to a family of psychology capabilities. These are proposed functions, not an existing service.
Clinics, electronic health-record companies, therapy platforms, mental health applications, research groups, employee-assistance providers and other software businesses.
These are possible buyer groups. Demand, integration needs and willingness to pay remain open commercial questions.
Usage-based charges, including per-token billing, alongside enterprise support or deployment arrangements. A token is a small piece of text processed or generated by a model.
Margins must account for inference, hosting, clinical evaluation, data acquisition, training, security and support. Downloadable weights do not mean free operation.
The potential advantage is cumulative. Owning a fine-tuned file alone does not establish a durable business. Rights in the base model, the data and the resulting work must also be clear. The source memo’s “you own” language should therefore be read as potential company assets subject to the relevant agreements.
Generating a summary and recommending treatment can carry different intended purposes and risks. UK MHRA guidance considers both purpose and functionality when determining medical-device status. Supplying an API or keeping a clinician in the loop does not automatically remove obligations. [9]
Define the roles of the model supplier and downstream application provider. Evidence, monitoring, privacy and responsibility for escalation need to fit the actual deployment and jurisdiction. No specific UK, EU or US classification has been established for this concept.
These are questions about the overall model programme, rather than a prescribed first product.
Compare continued pretraining plus fine-tuning against fine-tuning alone and strong general models with the same references and tools.
Test the contribution of supervision, preferences, literature and longitudinal information. The memo’s emphasis on clinician reasoning is a hypothesis worth measuring.
Build rubrics that preserve uncertainty and disagreement while detecting serious failures. Establish where automated judging helps and where it needs human review.
Research institutions, training organisations and clinical partners could contribute expertise and governed data. A credible clinical and machine-learning team is central to the programme.
Investigate transfer across settings, languages and populations, plus deployment cost and customer demand. A strong internal benchmark is one part of that answer.
The long-term vision is a clinical learning system:
one that connects expertise, evidence and change over time.
Open a term whenever the technical language gets in the way.
The numerical parameters learned during training. They influence how the model processes inputs and generates outputs.
A release that lets people access model parameters. Use, modification and commercial rights depend on its licence.
Further training of an existing model on a chosen body of material, often using next-token prediction.
Supervised fine-tuning: learning from examples that pair an instruction with a desired response.
Direct Preference Optimisation: training with preferred and rejected answers to shape behaviour.
Reinforcement learning using feedback from people / AI. The source of feedback and the training method are separate design choices.
A smaller trainable addition to a model. LoRA is one method for adapting a model without updating all its original weights.
Looking up relevant material and providing it to the model when answering, rather than relying only on learned parameters.
The surrounding software that controls workflow, permissions, tools and checks.
An embedding helps represent meaning for search. A reranker reorders retrieved results by relevance.
Testing on cases excluded from training. It estimates performance beyond the examples used to teach the model.
An API lets software request a capability. Inference is the process of running a trained model to produce an output.
Flathers, Torous and colleagues, August 2026. Benchmark methodology, scores and limitations.
The live repository describes a newer release. The comparison here preserves the supplied screenshot and paper v1.
Moonshot’s official terms, including the Model as a Service condition.
Official Modified MIT licence for K2.6. Separate from K3.
Rafailov and colleagues. The method behind learning directly from preferred and rejected responses.
Anthropic’s research on training with principles and AI feedback.
Google’s description of medical adaptation, intended use and limitations.
An example of a governed clinical-notes resource. Its existence does not establish rights for this project.
UK guidance on intended purpose, functionality and medical-device qualification.
UK guidance covering personal health information and related protections.
Hu and colleagues. Efficient adaptation with a smaller set of trainable parameters.
Gururangan and colleagues. Research on adapting language models to domains and tasks, not evidence for this proposed clinical system.
The main structure comes from the supplied ChatGPT concept document on psychology model training. The Kim × Flo meeting notes inform the lab ambition, the discussion of training versus runtime controls and the need for credible research partners.
The architecture, dataset priorities, model names such as PsychologyReward, potential customers and commercial advantage are proposals. They are not results from an implemented system. Method references establish what techniques do, not that this particular lab approach will succeed.
Benchmark figures preserve the supplied paper v1. Kimi terms were reviewed from official licence files on 28 September 2026. Other linked sources support method explanations and regulatory context. The page is self-contained and works offline; opening external sources requires an internet connection.