Reproduce “From Tags to Trees: Structuring Fine-Grained Knowledge for Controllable Data Selection in LLM Instruction Tuning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zihan Niu, Wenping Hu, Junmin Chen, Xiyue Wang, Tong Xu, Ruiming Tang
The TAGS framework structures fine-grained knowledge into a hierarchical tree for controllable data selection in Large Language Model instruction tuning. This method allows models to outperform those trained on full datasets by achieving up to a +5.84% average score improvement and a +5.38% gain in human-preference alignment while utilizing only 5% of the original data, and enables targeted knowledge-aligned sampling.
Key insights
- Explicitly structuring fine-grained knowledge into a hierarchical tree provides a more nuanced understanding of instruction data, crucial for precise data valuation and robust diversity assurance.
- A controllable sampling mechanism, guided by a knowledge tree and supported by distribution alignment, allows for fine-grained steering of LLM capabilities towards specific domains or task types.
- Significant data efficiency can be achieved by prioritizing structured, high-quality knowledge selection over raw data volume, leading to improved performance with substantially smaller instruction tuning datasets.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.