Reproduce “On Representation Redundancy in Large-Scale Instruction Tuning Data Selection” on OpenResearch

About this paper

Youwei Shu, Shaomian Zheng, Dingnan Jin, Wenjie Qu, Ziyao Guo, Qing Cui, Jun Zhou, Jiaheng Zhang

Researchers from NUS and Ant Group developed the Compressed Representation Data Selection (CRDS) framework, which mitigates redundancy in LLM semantic embeddings to improve instruction tuning data selection. CRDS-W, a variant of the framework, achieved higher average accuracy on benchmarks like GSM8K and MBPP, surpassing full-data baselines by up to 0.71% using only 3.5% of the training data.

Key insights

• LLM-generated semantic embeddings exhibit substantial redundancy, empirically confirmed by a pilot binarization experiment that yielded similar selection results with significantly reduced information.
• A carefully curated, smaller dataset (e.g., 3.5-5% of total data) selected using refined representations can lead to better LLM performance than training on the full, uncurated dataset.
• The effectiveness of an LLM encoder for data selection is not solely dependent on its size or recency; explicit representation refinement methods like CRDS are crucial.
• The transferability of CRDS methods might be limited for smaller LLMs, suggesting scale-dependent characteristics in their semantic representations.

Summary by alphaXiv.