Jun 2, 2025 · View a PDF of the paper titled T-SHIRT: Token-Selective Hierarchical Data Selection for Instruction Tuning, by Yanjun Fu and 2 other authors

We demonstrate that models instruction-tuned on a curated dataset (only 5% of the original size) using T-SHIRT can outperform those trained on the entire large-scale dataset by up to 5.48 …

In this paper, we present Token-Selective Hierarchical Data Selection for Instruction Tuning (T- SHIRT), a framework for selecting high-quality subsets from instruction tuning datasets.

We demonstrate that models instruction-tuned on a curated dataset (only 5% of the original size) using T-SHIRT can outperform those trained on the entire large-scale dataset by up to 5.48 …

Jun 4, 2025 · 简介：这篇论文提出了StochasTok，一个随机分割token的分词方案 diffusion LLM arxiv.org/pdf/2506.0041 标题： Accelerating Diffusion LLMs via Adaptive Parallel Decoding 关 …

Dec 16, 2025 · His research interests include LLMs, Efficient Post-training, and Trustworthy ML. (more) T-SHIRT: Token-Selective Hierarchical Data Selection for Instruction Tuning. Yanjun …

Jun 3, 2025 · In this paper, we present Token-Selective Hierarchical Data Selection for Instruction Tuning (T-Shirt), a framework for selecting high-quality subsets from instruction tuning datasets.

A new data selection framework, T-SHIRT, improves instruction tuning for Large Language Models by focusing on token-level informativeness and robust sample quality, leading to better …

Jun 2, 2025 · We demonstrate that models instruction-tuned on a curated dataset (only 5% of the original size) using T-SHIRT can outperform those trained on the entire large-scale dataset by …

Fu , F. Hamman, S. Dutta, "T-Shirt: Token-Selective Hierarchical Data Selection for Instruction Tuning" Neural Information Processing Systems (NeurIPS 2025). [Full-Paper]