| 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 | |