Sign in to view source links and access this dataset
Description
64,000 preference pairs from the Ultrafeedback dataset, sampled and ranked into difficulty levels. This dataset was created by ServiceNow-AI for curriculum learning experiments in model alignment, as described in the Curri-DPO paper. It was last updated on November 13, 2024.
Use Cases
Training language model alignment via Direct Preference Optimization (DPO) based on the provided preference pairs.
Implementing curriculum learning strategies for AI training based on the easy, medium, and hard difficulty levels.
Benchmarking model performance improvements using ranked response data.
Strengths
64,000 preference pairs provide a substantial base for training.
Pairs are organized into three curriculum difficulty levels (easy, medium, hard) based on response scores.
Explicitly created for and used in published research (Curri-DPO paper).
Limitations
Column-level documentation is absent; field semantics must be inferred after download.
Row count for the specific curated subset is unknown, which may limit suitability assessment.
Description metadata is limited; actual data quality requires manual inspection after download.
Provenance
Source
ServiceNow-AI, derived from the Ultrafeedback dataset.
Collection Method
Responses were ranked and organized into difficulty levels based on scores.
Freshness
Last updated 2024-11-13 09:22:09.
License is unknown; users must verify terms before use.