Hard 100

The 100 hardest GSM8K training problems (5 to 9 calculation steps), refined by 10 model families with short step-by-step reasoning. Accepted answers become training data; outvoted answers become the "rejected" side of preference pairs.

connecting…
Input
100
hard problems
Accepted
–
kept as training data
Rejected
–
no agreement, left out
Deciding
–
votes still coming in
Accepted and correct
match the official answer
Preference pairs
–
accepted vs outvoted answer

Progress

AcceptedRejectedDeciding

Which model and instruction got it right

Share of votes matching the official answer, per model and per instruction style. Grey = fewer than 3 votes.

Waiting for votes…

Nodes

Each node runs two models; its votes and status, live.

NodeStatusVotesAgreed
No nodes yet.

Live log

Every vote as it arrives: model, instruction, node and time taken. Answers are hidden.

  • Waiting for votes…

Latest decisions

Agreement = votes that match the network's answer. The official answer is never shown to the models.

#ProblemStepsAgreementNetworkOfficialResult
Waiting for the first decisions…