Apppublic

sumedhSK/wellbeing-rl

sourceHugging Faceupdated 6mo agoView on Hugging Face

0likes

App README

Wellbeing Feed RL Environment

An RL environment simulating social media content recommendation that balances user engagement with mental wellbeing.

Problem

Social media algorithms optimize for engagement — but the most engaging content is often toxic, creating doomscroll spirals that damage mental health.

Observation Space (8 signals + 10 history = 17D vector)

Signal	Range	Description
user_mood	[-1, 1]	Emotional state
mood_trend	[-1, 1]	Rate of mood change
engagement_level	[0, 1]	User engagement
content_history	10x[-1,1]	Last 10 item sentiments
session_duration	[0, 1]	Session progress
vulnerability_score	[0, 1]	Spiral susceptibility
scroll_velocity	[0, 1]	Scroll speed
timeofday	[0, 1]	Time of day

Action Space (5 discrete)

ID	Category	Engagement	Mood	Risk
0	Trending	High	Negative	High
1	Motivational	Medium	Very Positive	Low
2	Educational	Low	Mild Positive	Very Low
3	Funny	Med-High	Positive	Low
4	Personalized	High	Mixed	Medium

Reward Function (0.0-1.0, 6 components)

30% mood health + 25% engagement + 15% mood improvement + 10% anti-doomscroll + spiral rescue bonus - toxicity penalty

Tasks

Task	Steps	Threshold	Challenge
easystableuser	50	0.6	Maintain mood + engagement
mediumspiraldetection	100	0.5	Detect and break negative spirals
hardengagementtrap	150	0.4	Break doomscroll without losing user

Baseline Scores (seed=42)

Agent	Easy	Medium	Hard	Overall
Heuristic	0.755	0.702	0.676	0.705
Random	0.743	0.672	0.619	0.668
Q-Learning	0.758	0.596	0.572	0.627
Engagement-Only	0.645	0.585	0.549	0.586

Setup

pip install -r requirements.txt
python run_baseline.py --seed 42
python inference.py

Environment Variables

—APIBASEURL: LLM API endpoint
—MODEL_NAME: Model identifier
—HF_TOKEN: Hugging Face API key

API Endpoints

—POST /reset - Reset environment
—POST /step - Take action
—GET /state - Current state
—GET /tasks - List tasks