Trustless & fast-improving AI models — powered by SN74 on Gittensor.
We build open infrastructure for verified AI improvement: frontier teachers propose, confidential-compute GPUs prove, and deterministic evaluation rewards only measurable wins. No maintainer trust, no hand-waved datasets, no checkpoint-only claims.
| Repository | Role |
|---|---|
| SparkDistill | Open miner economy for Triton-native AI: verified datasets, training recipes, eval harness, and SN74 rewards on Gittensor. |
| SparkProof | Blackwell- and Hopper-attested Triton dataset generation — pinned frontier teachers, GPU compile/execute validation, Merkle proofs, and CPU-verifiable bundles for SparkDistill. |
- SparkProof miners generate stratified Triton trajectories on NVIDIA Blackwell (RTX PRO 6000) or Hopper (H100/H200) GPUs with Targon confidential compute + Intel TDX (Bittensor SN4), seal
sparkproof-2bundles, and publish verified datasets. - SparkDistill rewards two tracks: a dataset track that aggregates registry-approved miner datasets, and a training track that gates recipe/hyperparameter improvements against the pinned canonical mix — both cryptographically verified and scored by a deterministic harness so SN74 rewards marginal quality over the current frontier.
- Anyone can re-verify proofs from published Hugging Face bundles on ordinary CPU hardware — GPU + TDX attestation, claim hashes, and CPU-recomputable eval samples all travel with the data, so a full submission can be checked with no GPU at all.
Moving fast is key to success.
