love-mimi / Affine-kj01

Affine

Mine open reasoning.

Affine Discord

Introduction

Affine is an incentivized RL environment which pays miners which make incremental improvements on a set of tasks (for instance, program abduction or coding). The mechanism is sybil-proof (you can't cheat by deploying multiple miners), decoy-proof (you can't cheat by packing models into certain environments), copy-proof (you can't cheat by stealing models), overfitting-proof (you can't cheat by overfitting to a single env).

How does Affine work? Affine validators incentivize miners to submit models to Subnet 64 on Bittensor (a.k.a Chutes) where they are inference load balanced and publicly available. These models are evaluated on a set of RL-environments with validators looking for the model which dominates the pareto frontier -- namely the model which outcompetes all other models on all envs (see af validator ) The network is winners-take-all where miners are forced to copy, download and improve the pareto frontier model.

Why affine? Directed incentives for RL have never been achieved. The ability to direct intelligence and aggregate the work-effort of a large non-permissioned group of individuals on RL tasks will unlock fast advancement in intelligence, we intend to commoditize reasoning (intelligence's highest form) and break the intelligence sound barrier.

Installation

# Install uv Astral
curl -LsSf https://astral.sh/uv/install.sh | sh

# Clone and install Affine
git clone https://github.com/AffineFoundation/affine.git
cd affine
uv venv && source .venv/bin/activate && uv pip install -e .

# Verify installation
af

Validating

Set env vars, chutes api key.

# Copy .env and fill out validator items
cp .env.example .env

(Recommended): Run the validator with docker and watchtower autoupdate.

# Run the validator with watchtower.
docker-compose down && docker-compose pull && docker-compose up -d && docker-compose logs -f

Run the validator using the local override (build local image) + base compose

docker compose -f docker-compose.yml -f docker-compose.local.yml down --remove-orphans
docker compose -f docker-compose.yml -f docker-compose.local.yml up -d --build --remove-orphans
docker compose -f docker-compose.yml -f docker-compose.local.yml logs -f

Run the validator locally

# Start the validator with debug.
af -vv validate

Mining

IMPORTANT: you require a developer enabled account on Chutes to mine. Normal API keys cannot deploy chutes right now.

1. Set env vars.

# Copy .env and fill out validator items
cp .env.example .env

1. Miners need a chutes developer account ( chutes.ai )

chutes register

1. Register your miner to Affine (S120).

btcli subnet register --wallet.name <your cold> --wallet.hotkey <your hot>

1. Pull a model off the network.

af -vvv pull <uid to pull> --model_path <i.e. ./my_model>

1. Improve the model

... magic RL stuff ...

1. Push the model to your miner.

af -vvv push  --coldkey <your cold> --hotkey <your hot> --model_path <i.e. ./my_model>

SDK

Affine is also an SDK you can use to generate and evaluate models envs.

import affine as af

# Optionally turn on logging 
af.trace(); af.debug(); af.info()

# Get all miner info or only for UID =5
miners = await af.get_miners()
miner = await af.get_miners( 5 )

# Generate a SAT challenge
chal = await af.SAT.generate() 

# Generate a bunch.
chals = await af.ABDUCTION().many( 10 )
chals = await af.DEDUCTION().many( 10 )

# Query the model directly.
# NOTE: A CHUTES_API_KEY .env value is required for this command.
response = await af.query( chal.prompt, model = miner.model )

# Evaluate the response
evaluation = chal.evaluate( response ) 
print( evaluation.score )

# Async generator of results from last 100 blocks.
async for res in af.rollouts(100):
    print (res)          # Result objects