SHP-2: 4.8 Million Human Preferences Across 129 Subject Areas

4.8 million collective human preferences compare the helpfulness of two responses to questions or instructions. The dataset spans 129 diverse subject areas, from cooking to legal advice, and is an extended version of the original 385K SHP dataset. Created by stanfordnlp and updated in January 2024, it is intended for training RLHF reward models and NLG evaluation models.

Use Cases

• Training reward models for RLHF based on pairwise human preference judgments.
• Evaluating natural language generation models on the dimension of helpfulness.
• Analyzing human preference patterns across different domains like cooking and legal advice.
• Benchmarking alignment techniques for instruction-following AI systems.

Strengths

• Contains 4.8 million preference judgments, a significant scale for model training.
• Covers 129 distinct subject areas, suggesting broad topical diversity.
• Explicitly designed for RLHF and NLG evaluation, providing a clear purpose.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is known, but the distribution of preferences across the 129 subjects is unspecified.
• Data may reflect the biases inherent to the collection platform and participant pool.

Provenance

Source

stanfordnlp

Collection Method

Collection of human preferences over responses to questions/instructions.

Freshness

Last updated 2024-01-11 02:10:31