name: LvW-bench
description: >
  A local benchmark for tiny models. System 1: typed decisions (choice, score, noul)
  answered as probability distributions and scored on accuracy, KL from gold, Brier and ECE.
  System 2: generated answers on permissive subsets of GSM8K, MMLU, MMLU-Pro, TruthfulQA and
  IFEval plus a seeded generated set, once with thinking off (answer) and once with thinking
  on (reason).

leaderboard:
  base_models_only: false

tasks:
  - id: decide_accuracy
    config: decide_all
    split: test
  - id: decide_kl
    config: decide_all
    split: test
    lower_is_better: true
  - id: decide_brier
    config: decide_all
    split: test
    lower_is_better: true
  - id: decide_ece
    config: decide_all
    split: test
    lower_is_better: true
  - id: answer_accuracy
    config: generative_all
    split: test
  - id: reason_accuracy
    config: reason_subset
    split: test