Decloud Trainer Kit

Train models and earn rewards on Solana federated learning network.

Features

• 🔐 Wallet + Lighthouse IPFS integration
• 📡 WebSocket real-time round detection
• 🏋️ Automatic training and submission
• ⭐ Trainer rating system
• 💰 Easy reward claiming
• ⚙️ Configurable training parameters

Installation

cd decloud-trainer-kit

# Virtual environment (recommended)
python -m venv venv
source venv/bin/activate  # Linux/Mac
# venv\Scripts\activate   # Windows

# Install
pip install -r requirements.txt
pip install -e .

Quick Start

1. Setup

decloud-trainer setup

Prompts for:

• Solana wallet private key
• Network (devnet/mainnet)
• Training parameters

Lighthouse Storage API key is created automatically from your Solana wallet.

2. Create Trainer Profile

# Required before training - creates on-chain profile
decloud-trainer create-profile

# View your profile
decloud-trainer profile

3. Configure Datasets

# Set path for each dataset you want to train on
decloud-trainer dataset set Cifar10 /path/to/cifar10_data
decloud-trainer dataset set Mnist /path/to/mnist_data

# View configured datasets
decloud-trainer dataset list

4. Start Training

decloud-trainer start

Commands

Setup & Config

Command	Description
setup	Interactive setup wizard
status	Show trainer status
balance	Show wallet balance
network -n <net>	Change network

Profile Management

Command	Description
create-profile	Create trainer profile (required)
profile	Show your profile and rating

Dataset Configuration

Command	Description
dataset list	Show configured datasets
dataset set <n> <path>	Set dataset path
dataset remove <n>	Remove dataset
dataset available	Show all datasets

Training Settings

Command	Description
settings show	Show current settings
settings set min_reward 0.05	Set minimum reward
settings set epochs 10	Set training epochs
settings set batch_size 64	Set batch size
settings set lr 0.0001	Set learning rate

Training

Command	Description
start	Start auto-training (WebSocket)
rounds	Show active rounds
train <round_id>	Manually train for round
info <round_id>	Show round details

Rewards

Command	Description
claim <round_id>	Claim reward

Trainer Rating System

New trainers start at 5.00 ★ rating.

• If your training makes the model worse (post accuracy < pre accuracy), your rating is slashed by 0.01 ★
• Creators can set minimum rating requirements for their rounds
• Maintain quality training to keep your rating high

Check if you can participate in a round:

decloud-trainer info <round_id>
# Shows min rating requirement vs your rating

Dataset Format

Your local dataset should contain numpy arrays:

my_dataset/
├── embeddings_train.npy   # (N, embedding_dim) float32
├── labels_train.npy       # (N,) int64
├── embeddings_test.npy    # (M, embedding_dim) float32
└── labels_test.npy        # (M,) int64

Alternative naming:

• X_train.npy, y_train.npy
• train/embeddings.npy, train/labels.npy

How It Works

┌─────────────────────────────────────────────────────────────┐
│                    TRAINING FLOW                            │
└─────────────────────────────────────────────────────────────┘

1. WebSocket detects RoundCreated event
   ↓
2. Check: reward >= min_reward? dataset configured? rating OK?
   ↓
3. Download base model from IPFS
   ↓
4. Load YOUR local training data
   ↓
5. Train head model (fine-tune)
   ↓
6. Upload trained model to Lighthouse/IPFS
   ↓
7. Submit gradient CID to blockchain
   ↓
8. Wait for validators + round finalization
   ↓
9. Claim reward! (rating updated based on performance)

Configuration

Config file: ~/.decloud-trainer/config.json

{
  "private_key": "...",
  "network": "devnet",
  "lighthouse_api_key": "...",
  "min_reward": 0.01,
  "training_epochs": 5,
  "training_batch_size": 32,
  "learning_rate": 0.001,
  "dataset_paths": {
    "Cifar10": "/home/user/data/cifar10",
    "Mnist": "/home/user/data/mnist"
  }
}

Lighthouse Storage

Lighthouse API key is created automatically during setup using your Solana wallet for authentication. No separate account needed!

Tips

• Create profile first! You can't submit gradients without a profile
• Use GPU for faster training: CUDA_VISIBLE_DEVICES=0
• Set higher min_reward to only train for profitable rounds
• Keep embeddings small (~1000 test samples) for faster upload
• Maintain quality training to preserve your rating

License

MIT