Research map / Large language models / Instruction and fine-tuning
A Critical Look at Targeted Instruction Selection: Disentangling What Matters (and What Doesn’t)
Nihal V. Nayak, Paula Rodriguez-Diaz, Neha Hulkund, Sara Beery, David Alvarez-Melis
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use LESS gradient-based representation instead of RDS+ or EMBED for greedy round-robin selection (ablation) | Llama-2-7B, Tulu V2 pool, round-robin selection · BBH, Codex, GSM8K, TyDiQA, MMLU-Pro query loss | query-set cross-entropy loss | Improved |
| use LESS representation: subset-query distance quantile vs performance (ablation) | Llama-2-7B, round-robin selection · five target tasks, 10 distance quantiles | query loss and downstream performance correlation with distance | Improved |
| use RDS+ or EMBED representations (distance to query as predictor) (ablation) | Llama-2-7B, round-robin selection · five target tasks, 10 distance quantiles | Spearman correlation of distance with query loss | No clear effect |
| lower query loss from LESS representation (ablation) | LESS (RR) vs EMBED (RR) · first distance quantile, five target tasks | downstream performance | Mixed |
| LESS (RR) vs other representations across budgets (ablation) | Llama-2-7B, Tulu V2 pool · downstream tasks, budgets 500 to 10000 | task accuracy | Mixed |
| targeted selection (RDS+/EMBED) instead of random sampling (ablation) | Llama-2-7B · larger budgets, TyDiQA and MMLU-Pro | downstream performance | No clear effect |
| greedy round-robin selection at small budgets (ablation) | LESS, Llama-2-7B · low budgets, five target tasks | downstream performance | Improved |
| use UOT or KNN-KDE instead of round-robin at large budgets (ablation) | LESS, Llama-2-7B · high budgets, five target tasks | downstream performance | Improved |
| use doubly greedy (DG) selection instead of other algorithms (ablation) | LESS, Llama-2-7B · three of five datasets | downstream performance and query loss | Hurt |
| use UOT selection on LESS representations (ablation) | LESS, Llama-2-7B · five target tasks | query loss | Improved |
| switch candidate pool from Tulu V2 to Dolci Instruct (ablation) | targeted selection methods · downstream tasks | downstream performance | Mixed |
| increase selection budget (ablation) | LESS variants vs random · MMLU-Pro | loss gap to random sampling | Hurt |
Closest papers by meaning
- LESS: Selecting Influential Data for Targeted Instruction Tuning (ICML 2024 · similarity 0.93)
- What Makes Good Data for Alignment? A Comprehensive Study of Automatic Data Selection in Instruction Tuning (ICLR 2024 · similarity 0.91)
- SelectIT: Selective Instruction Tuning for LLMs via Uncertainty-Aware Self-Reflection (NeurIPS 2024 · similarity 0.91)
- Beyond IID: Optimizing Instruction Finetuning from the Perspective of Instruction Interaction and Dependency (AAAI 2025 · similarity 0.90)
- NICE Data Selection for Instruction Tuning in LLMs with Non-differentiable Evaluation Metric (ICML 2025 · similarity 0.90)
- PandaLM: An Automatic Evaluation Benchmark for LLM Instruction Tuning Optimization (ICLR 2024 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
