R-Select-100k: A Compact Instruction-Tuning Dataset for LLMs

R-Select-100k is a dataset of 100,000 instruction-tuning samples selected from a pool of over 3.4 million. It was created by author GX-XinGao using the R-Select multi-metric data selection framework and was last updated on June 17, 2026. The dataset is designed for fine-tuning large language models.

Use Cases

• Fine-tuning large language models based on a curated set of high-value instruction samples.
• Benchmarking data selection algorithms against the R-Select framework's 30-metric optimization.
• Studying the impact of dataset size on model performance, as suggested by the description's reference to 'Dataset Performance vs. dataset size'.

Strengths

• Contains 100,000 samples selected from a source pool of over 3.4 million.
• Selection process optimized 30 data-quality metrics using hierarchical Bayesian optimization.
• Framework (R-Select) is accepted by the KDD 2026 conference.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Description metadata is limited; actual data quality requires manual inspection after download.
• The source and nature of the original 3.4M sample pool are not detailed, which may obscure potential biases.

Provenance

Source

huggingface, author GX-XinGao

Collection Method

Curated from a heterogeneous instruction-tuning pool using the R-Select multi-metric selection framework.

Freshness

Last updated 2026-06-17 08:44:38; freshness should be verified.