name: Russian Olympiad Math Benchmark (ROMB)
description: >-
  Russian-language school olympiad mathematics, 2552 tasks. ROMB v2 measures
  single-attempt typed-JSON accuracy as a fraction in [0, 1]. Greedy decoding,
  thinking disabled, 1024 output tokens, no tools. Evaluation runs through
  Inspect AI with the custom ROMB scorer provided at
  https://huggingface.co/spaces/d0rj/romb-leaderboard.
  See README for the reproducible local runner and scoring protocol.
# Registration metadata only. HF's built-in inspect-ai YAML recipe validator
# rejects custom solvers/scorers. Omitting evaluation_framework is accepted by
# the Hub schema; it does not change the actual evaluator or imply activation.
tasks:
  - id: romb-v2-json-greedy-1024
    config: default
    split: test