Spaces:
Running
Running
Add MoodNet (2nd tiny model): 40 affect-anchored moods drive speech + behavior; energy/dream loop
b828eb7 A newer version of the Gradio SDK is available: 6.20.0
metadata
title: Dreamling
emoji: 🌱
colorFrom: indigo
colorTo: pink
sdk: gradio
sdk_version: 6.17.3
python_version: '3.12'
app_file: app.py
pinned: false
short_description: Raise a from-scratch baby AI that learns to talk.
🌱 Dreamling
A whimsical creature with a brain trained from scratch — about 0.3M parameters, no pretrained model, learning to talk live on a CPU. Fully off-grid.
- It's a newborn: at first it can only babble
dream? dream! dream-dream… - Talk to it and it learns your words — babble → letter-soup → fragments → speech.
- Tell it it's good 👍 or bad 👎 and you shape its personality: kindness makes it bold and bright 🦋, harshness makes it withdraw 🐚, spoiling it makes it wild 🐉.
- A live personality chart and its evolution reflect who it's becoming.
How it works — two tiny models from scratch
- Voice — a char-level Transformer (~0.3M params) seeded only on the word dream. While awake it babbles with what it knows and remembers what you say; when it dreams 🌙 it trains on the day's words (next-character learning) and wakes up able to use them.
- Mood — a second tiny net (~few hundred params) reads the interaction (praise/scolding, words it didn't understand, pace, tiredness) and outputs a point in an emotional space (valence × arousal × dominance). ~40 named moods are anchors in that space; it learns a temperament over time.
The mood steers the voice — tone, length, and behavior: confused → it echoes your unknown
words back (sun? sun?); exhausted → it yawns and begs to dream; scolded → it withdraws;
spoiled → it gets wild. Personality and mood decide which creature it evolves into
(🥚→🐛→🦋 / 🐚 / 🐉).
No pretrained weights, no GPU, no external APIs. Built for the Build Small hackathon (Thousand Token Wood). Each visitor gets their own creature; it lives for the session.