---
title: Ai Training Simulator
emoji: 🧠
colorFrom: blue
colorTo: green
sdk: gradio
sdk_version: 6.16.0
python_version: '3.13'
app_file: app.py
pinned: false
license: mit
short_description: Idle/strategy game
hf_oauth: true
---
# AI Training Simulator

**Build Small Hackathon 2026 — Chapter Two (Adventure in Thousand Token Wood)**

A browser strategy/idle game where you manage a fake AI startup, all in a terminal-like interface.

## How it works

- Start with **$100**
- Buy upgrades (model size, dataset size/quality, architecture, hardware)
- Run **Train Model** - cost and reward are simulated based on your upgrade combination
- The chat interface uses FlameF0X/Qwen3-4B-Distilled-Claude-4.6 with huggingface inference. It's a CoT model.
  - Low quality → high temperature → incoherent word salad
  - Upgrade your stack → temperature drops → coherent output
- Unlock **Cloud Inference** for recurring revenue

## Badges targeted
- 🎨 Off-Brand — fully custom UI (no default Gradio look)

## Files

| File | Purpose |
|------|---------|
| `app.py` | Gradio app + action handler |
| `state.py` | Pure Python game logic (train, upgrade, cloud) |
| `game_config.py` | **All balance numbers** — edit this to tune the game |
| `requirements.txt` | Only `gradio` needed (model runs in browser) |

## Customising / balancing

Everything tunable is in `game_config.py`:
- `STARTING_MONEY` — how much the player starts with
- `UPGRADES` — levels, costs, quality scores per upgrade
- `BASE_TRAIN_COST` / multiplier tables — training run economics
- `BASE_REWARD` + `QUALITY_EXPONENT` — reward curve shape
- `REWARD_NOISE_PCT` — variance per run
- `CLOUD_*` constants — cloud inference economics
- `CHAT_TEMP_*` — how temperature maps to model quality

Author: Martico2432