Subnet 34: BitMind

BitMind is Bittensor Subnet 34. Public BitMind materials also describe it as GAS, the Generative Adversarial Subnet, where synthetic-media detectors and synthetic-media generators compete in a shared evaluation loop.

What BitMind Provides

The BitMind README frames GAS as a Bittensor subnet inspired by generative adversarial networks. Detectors improve by learning to identify synthetic media, while generators improve by producing synthetic media that is harder for detectors to catch.

This design makes the subnet a Bittensor-native competition between detection and generation. Detection miners submit models that classify real versus synthetic content across supported media types, while generation miners create synthetic media in response to validator challenges. Validator weights from those evaluations feed into Yuma Consensus.

Two Mining Tracks

The discriminative mining guide describes discriminative miners as submitting binary classifiers for image, video, and audio media. For each sample, a model returns a probability distribution over real and synthetic classes, and semisynthetic media is treated as synthetic for scoring.

The generative mining guide describes generative miners as services that create synthetic images or videos from validator prompts. Those outputs must pass validation checks, and stronger adversarial performance comes from creating valid media that fools discriminative miners.

The two tracks are linked rather than independent. Generators provide fresh synthetic examples that challenge detectors, and detector performance creates pressure for generators to produce more convincing but still valid media.

Miner and Validator Roles

Discriminative miners contribute detection models. The discriminative guide says those models are evaluated on cloud infrastructure, so a discriminative miner’s main contribution is the submitted model rather than hosting inference hardware.

Generative miners contribute synthetic-media generation capacity. The generative guide describes a base reward for passing data-validation checks, a multiplier for adversarial performance against discriminators, and a sample-volume component that rewards broader participation in evaluations.

Validators challenge and score both miner types. The README describes validators as checking generated media, evaluating detection models against real and synthetic media, and setting weights from those evaluations.

Evaluation Context

The incentive documentation describes discriminator scoring with sn34_score, which combines discrimination accuracy with calibration quality. This matters because a detector should not only separate real from synthetic media; it should also express confidence in a way that matches the observed outcomes.

The same incentive document describes benchmark data as a mix of real, synthetic, and semisynthetic content, with regularly refreshed GAS-Station data and holdout data used to discourage overfitting. At the subnet level, that makes BitMind a moving target: a model that only memorizes public benchmarks should be less useful than one that generalizes to newer and withheld media.

For generation, the incentive document combines validation pass rate with adversarial performance. That keeps the generator side from being rewarded only for producing any output; useful generator work has to pass validation while still challenging the detector side of the subnet.

On-Chain Identity

Live SN34 subnet data is available on TaoStats. The source-backed detector, generator, and incentive details in this article come from public BitMind materials rather than from live identity fields.

Relationship to Yuma Consensus

Subnet 34 uses Yuma Consensus to convert the adversarial-evaluation weight vectors that validators submit into the emission shares distributed to miners and validators within the subnet each tempo. The linked documentation describes how validator weight submissions are aggregated into consensus weights for each miner registered on the subnet.

In BitMind’s context, validators challenge and score both discriminative miners — which submit detection models — and generative miners — which produce synthetic media — and translate those evaluation scores into weight vectors for the subnet. The Emission documentation describes how those consensus weights determine each participant’s share of the subnet’s accumulated emission each tempo.

Development Stage Context

The Introduction to Bittensor describes subnet development as moving from localnet to testnet and then mainnet. For BitMind (SN34), that sequence changes how readers should interpret synthetic media detection and generation evaluation outcomes.

In localnet, BitMind-compatible miners and validators can be developed and tested in an isolated environment. Localnet detection scores and emission outcomes do not represent production subnet performance.

On testnet, BitMind-compatible detection and generation components can be exercised in a shared, non-production network. Testnet evaluations and validator scores are separate from mainnet subnet state.

On mainnet, BitMind (SN34) is the live production subnet where detection miners and generation miners compete around synthetic media and validators evaluate those contributions to determine real Bittensor emissions. The BitMind repository describes the Generative Adversarial Subnet mechanism that applies on the production network.

The Bittensor Networks reference separates mainnet, testnet, and localnet. A detection result or emission outcome from one environment should not be read as representing production subnet performance in another environment.

Reader Boundary

Subnet 34 BitMind should not be read as generic Bittensor subnet documentation, a single image classifier benchmark, or proof that one detection score covers every media type. It names one subnet’s generative-adversarial synthetic-media competition on netuid 34 (Understanding Subnets, Glossary: Netuid).

Discriminative and Generative Tracks Interact

BitMind’s discriminative mining guide describes detection models for image, video, and audio media, while the generative mining guide describes miners creating synthetic media from validator prompts (Discriminative Mining, Generative Mining).

The two tracks are linked because generators supply fresh synthetic examples that challenge detectors.

sn34_score Combines Accuracy With Calibration

The incentive documentation defines discriminator scoring with sn34_score, which combines discrimination accuracy and calibration quality (Incentive documentation).

Detector quality is therefore measured on both separation and confidence behavior rather than raw accuracy alone.

Validator Weights Still Flow Through Yuma Consensus

Subnet 34 uses Yuma Consensus to convert validator weight submissions into emission shares each tempo (Yuma Consensus, Emission).