Zeus

<p align="center">
  <img src="static/zeus-icon.png" alt="Zeus Logo" width="150"/>
</p>
<h1 align="center">SN18: Zeus Environmental Forecasting Subnet<br><small>Ørpheus AI</small></h1>


![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)

[Website](https://www.zeussubnet.com/) · [X](https://x.com/zeussubnet) · [LinkedIn](https://www.linkedin.com/company/orpheus-ai-nl) · [Discord](https://discord.com/invite/bittensor) · [huggingface](https://huggingface.co/orpheus-zeus)



## Quick Links

- [Mining Guide ⛏️](docs/Mining.md)
- [Incentive mechanism 🎁](docs/ScoringChallengesCalculatingWeights.ipynb)
- [Validator Guide 🔧](docs/Validating.md)

> [!IMPORTANT]
> If you are new to Bittensor, we recommend familiarizing yourself with the basics on the [Bittensor Website](https://bittensor.com/) before proceeding.



## Table of Contents

- [Who are we](#who-are-we)
- [What we do](#what-we-do)
  - [Overview](#overview)
  - [Purpose](#purpose)
  - [Features](#features)
- [Feedback and Contributions](#feedback-and-contributions)
- [License](#license)
- [Contacts](#contacts)



## Who are we

Ørpheus AI builds Zeus (Bittensor subnet 18): a decentralized weather forecasting network that turns competitive machine-learning forecasts into actionable atmospheric intelligence for energy markets. Find out more on our website!

## What we do
### Overview

The Zeus Subnet leverages advanced AI models within the Bittensor network to forecast environmental data on a decentralized, incentive-driven framework. The datasource for this subnet consists of ERA5 reanalysis data from the Climate Data Store (CDS) of the European Union's Earth observation programme (Copernicus). This comprises the largest global environmental dataset to date, containing hourly measurements from 1940 until the present across hundreds of variables. Validators issue global ERA5 forecasting challenges for four surface variables used heavily in energy trading: 2 m temperature, 100 m u- and v-components of wind, and surface solar radiation downwards. Miners compete to produce the best forecasts on the full Earth grid; validators score reveals against ERA5, keep rank history, and set subnet weights from verified performance.

### Purpose

Traditionally, environmental forecasting relies on physics-based numerical weather prediction (NWP). While this allows for very accurate predictions, it is also highly cost-ineffective, requiring large amounts of computing power for a single forecast. Furthermore, predictions are time expensive to obtain, since the simulation process of these NWP algorithms can take multiple hours to finish. Currently, there is a lot of ongoing research into the development of intelligent, data-driven algorithms for environmental prediction. Such algorithms can potentially be much faster, more accurate, at a fraction of the cost and carbon emissions. This subnet incentives the development of novel and groundbreaking architectures for environmental data prediction. Through the continuous evolution of this subnet, we are able to allow miners to tackle increasingly difficult problems over time.

### Features

- **Hourly 15-day forecasts for energy trading desks.** Zeus targets hourly forecasts out to 15 days for four surface variables that matter to power and gas books: 2 m temperature, 100 m eastward and northward wind, and surface solar radiation downwards. Challenges run on a global 0.25° ERA5 grid with both shorter and longer horizons so the market prices skill where desks actually use weather—not only at synoptic publish times of traditional NWP.
- **Dynamic attention shifting by client need.** Scoring and incentives can shift focus depending on commercial requirements—horizon, resolution, variables, and coordinates—through challenge weighting and geographic scalars (for example regional boosts over Europe and Germany). That lets pilots and products steer miner effort toward the regions and lead times that improve portfolio outcomes, rather than treating every grid cell as equally valuable.
- **Open-source verification of miner predictions via Hugging Face.** A public dataset and tutorial support trustless checks of forecast quality: see the [trustless verification tutorial](https://huggingface.co/datasets/orpheus-zeus/Zeus-API-forecasts/blob/main/trustless_verification_tutorial.ipynb). A selection of predictions has been available since 17 June, with a 7-day upload delay. Businesses can verify forecasts using the blockchain and the hashes stored in prediction metadata; the notebook walks through the end-to-end process.
- **Anti-gaming through on-chain commit-reveal.** Miners commit a hash of their compressed prediction to the blockchain before revealing the full forecast. That has two significant outcomes: it effectively stops relay mining on subnet 18 (copying another miner’s answer after seeing it), and it allows businesses to verify predictions as described above by matching revealed bytes to the on-chain commitment.
- **Epoch dynamics and dynamic burn.** Weight setting runs on a dedicated epoch schedule: a background setter refreshes burn data near epoch end, waits for the configured block window, respects chain weight rate limits, then emits weights from a fresh rank snapshot. **Dynamic burn** assigns a varying share of emissions to a burn UID each epoch (fetched from the performance API, with a configurable fallback). The goal is to deter **weight-copying validators**—operators that mirror another validator’s weights without doing the scoring work—by making copied weight vectors misaligned when burn changes every epoch, so honest validators that compute ranks and apply the current burn remain the ones emissions follow.
- **Incentive and validator infrastructure tied to verified skill.** Validators store challenge metadata, rank history, and per-challenge top miners in local SQLite databases. Subnet weights come from rolling averages of recent ranks (separate windows for short- and long-horizon challe