PPE-GPQA-Best-of-K: Preference Evaluation Set for Reward Models

lmarena-ai's PPE-GPQA-Best-of-K dataset contains a correctness preference evaluation set for Preference Proxy Evaluations. The prompts are sampled from the GPQA dataset, and the collection is intended for benchmarking and evaluation, not for training. The dataset was last updated on October 22, 2024.

Use Cases

• Benchmarking reward model performance based on correctness preferences.
• Evaluating preference proxy methods using sampled GPQA prompts.
• Comparing different reinforcement learning from human feedback (RLHF) techniques on a standardized test set.

Strengths

• Dataset is specifically designed for benchmarking and evaluation, providing a clear use case.
• Prompts are sampled from the established GPQA dataset, suggesting a connection to a known benchmark.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.

Provenance

Source

lmarena-ai on Hugging Face.

Collection Method

Prompts sampled from the GPQA dataset.

Freshness

Last updated 2024-10-22 08:00:01.