PaTaRM: Training Data for Pairwise and Pointwise Alignment

PaTaRM-data is the training data for the PaTaRM series, a collection for aligning language models. The dataset contains two subsets: one with 35.6k samples for supervised fine-tuning (SFT) and another with 41.7k samples for reinforcement learning (RL). It was created by AIJian and last updated on April 1, 2026.

Use Cases

• Supervised fine-tuning of language models based on the 35.6k SFT samples
• Training reinforcement learning from human feedback (RLHF) agents based on the 41.7k RL samples
• Bridging pairwise and pointwise alignment signals as referenced in the associated research paper

Strengths

• Contains 35.6k samples for supervised fine-tuning
• Contains 41.7k samples for reinforcement learning training
• Associated with a cited research paper (Jian, 2026)

Limitations

• Column-level documentation is absent; field semantics must be inferred after download
• Row count for the full dataset is unknown, which may limit suitability assessment
• Data format and sample data are unavailable for preview

Provenance

Source

AIJian

Collection Method

Likely created for training the PaTaRM series models, but specific gathering method is not detailed.

Freshness

Last updated 2026-04-01 06:19:25