Gradients

Gradients is Subnet 56 on Bittensor, maintained by Rayon Labs, and turns model fine-tuning into a training market that miners bid into. A user picks a model and dataset in the browser, the platform suggests column mappings, and the job is dispatched: miners submit a training script pinned to an exact commit, validators run it on dedicated GPUs and score the resulting model against hidden evaluation sets, with exponentially higher weights for the top entries. Finished models can be pulled from Hugging Face and tracked in Weights & Biases. An API at POST /v1/tasks/create accepts a model repository, dataset repository, instruction column and target completion hours.

Tournaments were layered on from 2025: each round runs four to seven days, elimination narrows the field, and the winning script is published to a public repository as the baseline for the next round. A separate path serves paying customers, who upload data while miners submit within a fixed window and are graded on held-out tests. The codebase kept a high commit rate through 2026; a 25 August change added tests for contestant selection and dataset propagation and addressed memory pressure during evaluation.