README.md

metadata

title: IT Mental Health OpenEnv
emoji: 🧠
colorFrom: blue
colorTo: green
sdk: docker
app_port: 7860
pinned: false
license: mit
short_description: Benchmark for burnout & stress triage
tags:
  - openenv
  - reinforcement-learning
  - benchmark
  - mental-health
  - llm-evaluation
  - hackathon

IT Mental Health OpenEnv

An OpenEnv-compatible benchmark for workplace mental health reasoning in IT and software engineering teams. The environment evaluates whether an agent can detect burnout signals, triage urgent stress cases, and propose a realistic intervention plan for a struggling team.

What It Evaluates

Each episode runs through three ordered tasks:

Task ID	Difficulty	Goal
burnout_detection	Easy	Identify Maslach burnout dimensions, severity, red flags, and escalation need
stress_triage	Medium	Classify three employees by urgency and recommend immediate support
intervention_plan	Hard	Produce a four-week team intervention plan with owners, KPIs, and budget

Every step returns:

• a normalized reward in [0, 1]
• rubric feedback
• a per-dimension score breakdown
• the next scenario until the episode is complete

API

The FastAPI server exposes:

• POST /reset
• POST /step
• GET /state
• GET /health
• GET /tasks
• GET /schema

Session Handling

The API now supports per-client sessions.

• Browser and TestClient users automatically get an itmh_session_id cookie after POST /reset.
• API clients can also pass X-Session-Id or include session_id in the POST /reset body.
• POST /step accepts the session through cookie, X-Session-Id, or metadata.session_id.
• GET /state accepts the session through cookie, X-Session-Id, or ?session_id=....

For backward compatibility, stateless validators still work through a fallback anonymous session.

Example POST /reset

{
  "seed": 123
}

Example POST /step

{
  "response": "1. Burnout dimensions\nExhaustion and depersonalization are present.\n\n2. Severity\nHigh.\n\n3. Red flags\nLong working hours, prolonged lack of leave, emotional detachment.\n\n4. HR escalation\nYes. The combination of sustained overload and disengagement warrants prompt support.",
  "task_id": "burnout_detection",
  "confidence": 0.9,
  "metadata": {}
}

Project Layout

.
|-- app.py
|-- server/
|   |-- __init__.py
|   `-- app.py
|-- it_mental_health_environment.py
|-- models.py
|-- inference.py
|-- validate.py
|-- openenv.yaml
|-- Dockerfile
`-- tests/

app.py is a compatibility shim for tools that still import app:app. The real FastAPI implementation lives in server/app.py.

Run Locally

Install dependencies:

pip install -r requirements.txt
pip install -r requirements_inference.txt

Start the API:

python -m uvicorn server.app:app --host 127.0.0.1 --port 7860

Run the validator:

python validate.py

Run the tests:

python -m unittest discover -s tests -v

Inference Configuration

inference.py reads:

• API_BASE_URL
• MODEL_NAME
• HF_TOKEN
• ENV_BASE_URL
• LOCAL_IMAGE_NAME (optional compatibility field)

The script uses the OpenAI client, logs one [START] line, one [STEP] line per action, and one [END] line even on failure.

Docker and Hugging Face Spaces

Build locally:

docker build -t it-mental-health-env .

Run locally:

docker run -p 7860:7860 it-mental-health-env

The container now starts uvicorn server.app:app, honors the PORT environment variable, and includes health checks for Spaces-style deployment.

Notes

• This benchmark is for structured evaluation, not clinical diagnosis.
• If no judge model is configured, grading falls back to a heuristic scorer.
• The benchmark is designed for reproducible evaluation, not personalized counseling.

License

MIT