Reproducible Evaluation With Fixed Question Set

via “benchmark reproducibility through fixed question sets and seed management”

Multi-turn conversation benchmark — 80 questions, 8 categories, GPT-4 as judge.

Unique: Treats reproducibility as a first-class concern by versioning questions, recording all inference parameters, and publishing metadata alongside results. Questions are public, enabling external verification.

vs others: More reproducible than proprietary benchmarks (which don't publish questions); more rigorous than informal evaluation practices that don't track parameters.

57-subject benchmark, the standard metric for comparing LLMs.

Unique: Immutable, versioned dataset published on Hugging Face ensures that any builder can download and evaluate against the exact same 15,908 questions used in published research. No question generation variance, sampling randomness, or dataset drift between evaluation runs.

vs others: More reproducible than dynamically-generated benchmarks or evaluation sets that vary between researchers; enables verification of published results and fair comparison across models and time periods.

via “fixed 2500-question snapshot for reproducibility”

Hardest exam questions from thousands of experts.

Unique: Decouples the fixed reference benchmark (2,500 questions, Nature publication, reproducible) from the rolling version (HLE-Rolling, community contributions, evolving). This dual-version approach allows researchers to use the stable snapshot for reproducible comparisons while the rolling version evolves with community input, balancing reproducibility and adaptability.

vs others: Provides reproducibility guarantees that rolling benchmarks (HELM) cannot offer, since HELM's question set changes over time. However, it sacrifices adaptability compared to rolling benchmarks, potentially becoming outdated as AI capabilities advance. The fixed snapshot is more reproducible than GitHub-based benchmarks without version pinning.