Subnet 56: Gradients
Gradients is Bittensor Subnet 56. The live Finney identity lists its subnet name as Gradients, its GitHub repository as gradients-ai/G.O.D, its subnet URL as gradients.io, and its description as “Best AutoML plaftorm in the world.”
The G.O.D repository describes the project as Gradients on Demand: a subnet for competitive LLM and diffusion model training, where AutoML methods compete in tournaments. Rather than serving a model endpoint, the subnet evaluates training approaches by having validators run miner-submitted training code on standardized infrastructure.
What Gradients Provides
Gradients provides a tournament loop for model-training methods. Its README describes competitive events where miners submit open-source training repositories, validators execute those repositories, and the system advances participants through rounds until tournament winners are selected.
The repository describes tournament categories for text, image, and environment-based tasks. At the article level, the subnet’s output is not one fixed model; it is a recurring evaluation process for training strategies that can improve LLM, diffusion, and environment-task performance. The resulting tournament scores inform subnet weights that flow through Yuma Consensus.
Miner and Validator Roles
Miners provide training methods as open-source repositories. The current miner guide says tournament miners expose repository details for the relevant task type, and the tournament system builds and runs that training code within defined time and resource limits.
Validators run the evaluation side of the tournament. The current README and miner guide say validators execute miner repositories on dedicated infrastructure, run the submitted training code, evaluate the resulting models, and use tournament results for subnet weighting. This is the subnet-specific form of the broader Bittensor mining and validating pattern described in the miner and validator documentation.
Tournament Entry Context
The current README and miner guide make the miner artifact more specific than “a model.” A Gradients miner submits the training method by exposing a repository and a fixed commit for the tournament type. Validators clone that submitted commit, build the training environment, run the training code on validator-managed GPUs, and then evaluate the resulting model against other tournament entries.
That means the competition is over reproducible training procedure, not just a finished model endpoint. The miner guide describes repository requirements, readable source code, required licensing files, and a final model output path because validators need enough structure to run each entry under the same tournament conditions. Winning repositories are later released through the public winners organization described in the README.
The tournament types also clarify the subnet’s scope. The README lists text, image, and environment tournaments on separate weekly schedules, while the miner guide describes independent requirements for each task type. Gradients therefore functions as a recurring AutoML tournament system: miners submit training code for the active task family, validators execute and compare those submissions, and tournament outcomes become the basis for subnet weights.
The miner guide also describes participant validation before a tournament activates. A pending tournament waits for enough validated miners in its task type, so the contest is not only scheduled by time; it also depends on having a qualified field of entries.
References: G.O.D README source, Gradients miner guide source
On-Chain Identity
Live SN56 data is available on TaoStats. Readers can also
reproduce the chain identity fields with the Bittensor CLI command documented in the
btcli reference:
btcli subnets get-identity --netuid 56 --network finney.
That Finney identity reports the subnet name as Gradients, the subnet URL as gradients.io, and the GitHub repo as gradients-ai/G.O.D.
Relationship to Yuma Consensus
Subnet 56 uses Yuma Consensus to convert the tournament-score 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 Gradients’s context, validators clone miner-submitted training repositories, execute the submitted training code on dedicated infrastructure within defined time and resource limits, evaluate the resulting models in competitive rounds, and translate tournament outcomes 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 Gradients (SN56), that sequence changes how readers should interpret AutoML tournament examples and training competition scoring outcomes.
In localnet, Gradients-compatible miners and validators can be developed and tested in an isolated environment. Localnet tournament results and emission outcomes do not represent production subnet performance.
On testnet, Gradients-compatible training submission and tournament evaluation workflows can be exercised in a shared, non-production network. Testnet tournament scores and validator weights are separate from mainnet subnet state.
On mainnet, Gradients (SN56) is the live production subnet where miners submit open-source training repositories and validators execute them on standardized infrastructure to determine real Bittensor emissions. The G.O.D repository is the registered project repository for SN56 on the production network.
The Bittensor Networks reference separates mainnet, testnet, and localnet. A tournament result or emission outcome from one environment should not be read as representing production subnet performance in another environment.
Reader Boundary
Subnet 56 Gradients should not be read as generic Bittensor subnet documentation, a hosted model inference API, or proof that one finished model checkpoint alone wins permanently. It names one subnet’s AutoML training-method tournament on netuid 56 (Understanding Subnets, Glossary: Netuid).
Fixed Commits Let Validators Run Submitted Training Code
The Gradients miner guide describes miners exposing a repository and fixed commit for validators to clone, build, and execute on validator-managed GPUs (Gradients miner guide source).
The competition therefore centers on reproducible training procedure rather than opaque endpoints.
Text, Image, and Environment Tournaments Run on Separate Schedules
The G.O.D README lists text, image, and environment tournament categories on separate weekly schedules (G.O.D README source).
Each task family keeps its own entry requirements and evaluation rhythm.
Validator Weights Still Flow Through Yuma Consensus
Subnet 56 uses Yuma Consensus to convert validator weight submissions into emission shares each tempo (Yuma Consensus, Emission).