PRISM

Architecture mining, training variants, scoring, and operators.

PRISMPRISM

Benchmarks and scaling

Benchmark families, resource strategy, sharding, caching, and scaling behavior for validator workers.

#prism/benchmarks-scaling
prismbenchmarksscalinggpu

Benchmark types

PRISM should combine architecture-sensitive tests with training-variant tests.

  • Heldout language or reasoning tasks.
  • Long-context tasks for context claims.
  • Stability runs across seeds or shards.
  • Efficiency checks for memory, latency, active parameters, and GPU hours.
  • Regression tasks that prevent overfitting to one benchmark family.

Scaling controls

Validators need predictable resource bounds so one submission cannot starve the queue.

ControlPurpose
GPU class labelsRoute jobs to compatible worker pools.
Wall-clock timeoutStop runaway builds or benchmarks.
Artifact cacheReuse verified artifacts by checksum.
Shard schedulerParallelize tasks without revealing hidden-set integrity.