CLaRa: Training Data for a Retrieval-Augmented Generation Model

CLaRa is a dataset for training and evaluating a model designed to bridge retrieval and generation with continuous latent reasoning. The data is organized into three main categories: pretraining, instruction tuning, and end-to-end tuning. It was created by Apple and last updated on the Hugging Face platform in December 2025.

Use Cases

• Pretraining a language model compressor based on the large-scale pretraining data.
• Instruction tuning a model for specific tasks using the instruction tuning data.
• End-to-end fine-tuning of a retrieval-augmented generation pipeline.
• Benchmarking model performance on tasks involving latent reasoning and retrieval.

Strengths

• Data is structured into three distinct training stages (pretraining, instruction tuning, end-to-end tuning), providing a clear development pipeline.
• Dataset is associated with a published research paper (arXiv:2511.18659), suggesting a defined research context.

Limitations

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

Provenance

Source

Apple

Collection Method

Created as the official dataset for the CLaRa research paper.

Freshness

Last updated 2025-12-16 19:34:29