Skip to content
← Bounties

Federated rounds

Two organizations train the same model on data neither one shares. Each sends only a weight update, each update is hashed, and the averaged model is evaluated on a held-out set both can see.

Rounds

  • Round 10.2158
    • org a · 3 episodes kept privateΔ f7e8dcf64379…
    • org b · 2 episodes kept privateΔ e06663f985ab…
    min 2 participants

    global 0x9173e6365ff9aacb30b2718e150fb2a99dfe4b325f75e6a13ae7d98fed3a662d

  • Round 20.2119best
    • org a · 3 episodes kept privateΔ 3c6a115670e9…
    • org b · 2 episodes kept privateΔ bc380728594f…
    min 2 participants

    global 0xa70c829b68bf2652b811a23bc5ae3fd1d245f9f19879ff1d6558e4743547d3b3

  • Round 30.2311
    • org a · 3 episodes kept privateΔ fd62bec804c6…
    • org b · 2 episodes kept privateΔ db85af9bcc59…
    min 2 participants

    global 0x1591649383152d0555e7996500f3009a851932965343e8d39b6e5f1758b522f8

  • Round 40.2634
    • org a · 3 episodes kept privateΔ d8360d83bb19…
    • org b · 2 episodes kept privateΔ 30d5a8f1b8a4…
    min 2 participants

    global 0xf01acf553603ff800db27b0f9b7c5ffb9e38f233b09dd2509cff540fb4ab5113

What the rounds show

Best round
0.2119
No-change baseline
0.1973
Shared model
692,544 params

Raw data never moved: each organization sent 692,544 numbers and a hash. That is the coordination the protocol needs, and it works.