Research map / Large language models / Instruction and fine-tuning
The Flan Collection: Designing Data and Methods for Effective Instruction Tuning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| add mixed zero-shot and few-shot prompt templates at training (ablation) | Flan-T5 XL (3B) · Held-In, MMLU | accuracy (few-shot): Held-In few-shot 62.2 -> 74.8; MMLU few-shot 38.7 -> 52.4 | Helped |
| add CoT training data (ablation) | Flan-T5 XL · MMLU and BBH-CoT | accuracy: MMLU 47.5/46.9 -> 50.3/52.4; BBH-CoT 18.2/12.0 -> 33.9/35.2 | Helped |
| add input inversion augmentation (ablation) | Flan-T5 XL · Held-Out MMLU and BBH | accuracy: MMLU 41.7/41.2 -> 50.3/52.4; BBH 18.4/24.2 -> 26.2/39.3 | Helped |
| add input inversion augmentation (ablation) | Flan-T5 XL · Held-In | accuracy: Held-In 73.8/74.8 vs 73.8/74.1 without | No clear effect |
| add mixture balancing of task sources (weighted vs equal) (ablation) | Flan-T5 XL · Held-In, CoT, MMLU, BBH | accuracy: MMLU 45.4/45.8 -> 50.3/52.4; Held-In 71.2/73.1 -> 73.8/74.8 | Helped |
| omit Flan 2021 from equally weighted mixture (ablation) | task mixture study · Held-In, CoT, MMLU | accuracy: MMLU 47.3 -> 45.7; Held-In 64.9 -> 55.3 | Hurt |
| Flan-T5 XL full recipe vs prior public collections on T5-XL | instruction tuning collection · Held-Out MMLU | accuracy: +4.2 zero-shot over next best T5-XL | Helped |
| Flan 2022 collection vs Super-Natural Instructions/P3++ on T5-XL in BBH | instruction tuning collection · BBH (zero-shot) | accuracy: -1.9 vs next best T5-XL (Flan 2021 28.1 vs 26.2) | Hurt |
Closest papers by meaning
- LESS: Selecting Influential Data for Targeted Instruction Tuning (ICML 2024 · similarity 0.89)
- What Makes Good Data for Alignment? A Comprehensive Study of Automatic Data Selection in Instruction Tuning (ICLR 2024 · similarity 0.88)
- Instruct-SkillMix: A Powerful Pipeline for LLM Instruction Tuning (ICLR 2025 · similarity 0.88)
- Adapt-$\infty$: Scalable Continual Multimodal Instruction Tuning via Dynamic Data Selection (ICLR 2025 · similarity 0.87)
- PandaLM: An Automatic Evaluation Benchmark for LLM Instruction Tuning Optimization (ICLR 2024 · similarity 0.87)
- Long Is More for Alignment: A Simple but Tough-to-Beat Baseline for Instruction Fine-Tuning (ICML 2024 · similarity 0.87)
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.