Angel-Observer
Angel-Observer reads a patient's presenting complaints and builds a directed symptom network (a Network Model): the thoughts, emotions, behaviors, bodily sensations and events in the case, and which ones set off which. It is stage 1 of Angel, a two-stage simulator of psychotherapy patients; Angel-Actor then role-plays the patient.
🎮 Live demo · 💻 Code: ANGEL-UserSim · 🎭 Angel-Actor
short description ──► Angel-Observer ──► long profile + symptom network ──► Angel-Actor ──► patient replies
What it does
One checkpoint answers two prompts:
| Step | Input | Output |
|---|---|---|
| S1: nodes | presenting complaints | <think>…</think><GRAPH>{"symptoms": [...], "external_factors": [...]}</GRAPH> |
| S2: edges | complaints + the S1 nodes | <think>…</think><GRAPH>{"links": [{"from": ..., "to": ...}]}</GRAPH> |
In the Angel pipeline it also expands a short patient description into a detailed long profile for the Actor.
How to use
The system prompts the model was trained on are in the code repository, so the easiest way is through it:
git clone https://github.com/Scarelette/ANGEL-UserSim.git && cd ANGEL-UserSim
pip install -r model_training/observer/requirements-observer.txt
# one symptom network per case (JSONL with a "Complaints" field)
python -m model_training.observer.predict_network \
--input data/examples/observer/case_reports.jsonl --output networks.jsonl
The model is downloaded from this page on first use. From Python:
from model_training.observer.predict_network import load_observer, predict_network
tokenizer, model = load_observer("ChengLi0228/Angel-Observer")
net = predict_network(tokenizer, model, "Alex, 29, reports three months of low mood, "
"poor sleep and withdrawing from friends after being laid off ...")
print(net["symptoms"])
print(net["graph"]) # [{"from": "being laid off", "to": "low mood"}, ...]
For the full Angel patient (Observer + Actor) see
model_usage/.
It uses the Qwen3 chat template, sampling with temperature 0.1 and top-p 0.9,
and a bf16 model needs about 17 GB of GPU memory.
Training
Qwen3-8B, fine-tuned in four stages (QLoRA adapters on a 4-bit copy, each merged back in bf16):
| Stage | Data | Settings |
|---|---|---|
| SFT, S1 (nodes) | 5,599 rows | LoRA r=64 / α=16, lr 1e-4, 3 epochs |
| GRPO, S1 | 5,089 prompts, 600 steps | reward 0.4 · format + 0.6 · node recall (MiniLM cosine ≥ 0.75 against reference nodes) |
| SFT, S2 (edges) | 2,138 rows | LoRA r=64 / α=16, lr 1e-4, 3 epochs |
| GRPO, S2 | 800 steps | reward 0.1 · format + 0.6 · edge precision + 0.3 · node coverage − size penalty; edges scored by a gpt-5-mini judge |
The training data were built with GPT-5 from 510 of the 516 published
psychotherapy case reports in PSYCHE, our graph-grounded dataset for
psychological user simulation (release coming soon). The case reports are
copyrighted and are not redistributed. Full recipe:
model_training/observer/.
Intended use and limitations
- Research use in user simulation, for example generating varied, structured patient cases to train or evaluate therapy-support systems and clinicians-in-training.
- Not a diagnostic or clinical tool. The networks are the model's reading of a text, not clinical assessments, and they can be wrong, incomplete or reflect biases in the source case reports.
- Trained on English case-report language; other languages and very different writing styles are untested.
- Case material can involve self-harm, suicidality, trauma and substance use, so the model's outputs can too.
Citation
Coming soon.