Harvard-DCML/tis-subset-datasets-Llama-2-7b-hf

Targeted Instruction Selection Subsets (Llama-2-7b-hf)

This repository contains pre-computed instruction training subsets selected from a large candidate pool for targeted instruction fine-tuning, as presented in the paper A Critical Look at Targeted Instruction Selection: Disentangling What Matters (and What Doesn't).

• Paper: https://huggingface.co/papers/2602.14696
• GitHub Repository: https://github.com/dcml-lab/targeted-instruction-selection

Description

Instruction fine-tuning of large language models (LLMs) often involves selecting a subset of instruction training data from a large candidate pool, using a small query set from the target task. This dataset provides pre-computed subsets for Llama-2-7b-hf generated using different data representations (RDS+, EMBED, and LESS) and selection algorithms (Round Robin, Doubly Greedy, UOT, KNN-Uniform, and KNN-KDE).

Subsets are provided for several target tasks, including: BBH, Codex, GSM8K, TyDiQA, and MMLU-Pro.

Dataset Structure

The repository contains multiple configurations based on different selection parameters. Config names follow the format: {representation}_{algorithm}_{query_set}_{budget}.

For example, less_rr_bbh_10000 refers to a subset selected using LESS representation with a round-robin algorithm targeting the BBH task with a budget of 10,000 samples.

Each sample includes:

• dataset: The source dataset name.
• id: A unique identifier for the example.
• messages: Conversation history in a role/content list format.
• index: The original index of the sample in the candidate pool.

Citation

If you find this work or the subsets useful, please consider citing the paper:

@article{nayak2026critical,
  title={A Critical Look at Targeted Instruction Selection: Disentangling What Matters (and What Doesn’t)},
  author={Nayak, Nihal V. and Rodriguez-Diaz, Paula and Hulkund, Neha and Beery, Sara and Alvarez-Melis, David},
  journal={arXiv},
  year={2026},
}