Research map / Large language models / Instruction and fine-tuning
Skill-it! A data-driven skills framework for understanding and training language models
Mayee F Chen, Nicholas Roberts, Kush Bhatia, Jue WANG, Ce Zhang, Frederic Sala, Christopher Re
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 Skill-it! instead of random sampling for continual pre-training | LEGO synthetic · LEGO synthetic, continual pre-training | accuracy: 36.5 points higher accuracy than random sampling | Improved |
| Use Skill-it! instead of random sampling for continual pre-training | Natural Instructions · Natural Instructions, continual pre-training | validation loss: 3.2% lower | Improved |
| Use Skill-it! instead of training only on the target skill | Natural Instructions Spanish question generation · Natural Instructions, Spanish question generation, fine-tuning | validation loss: 5.3% | Improved |
| Use Skill-it! instead of training only on the target skill | Natural Instructions Stance Detection · Natural Instructions, Stance Detection, fine-tuning | validation loss: 13.6% | Improved |
| Use Skill-it! instead of target-skill-only training | LEGO synthetic skill 3 · LEGO synthetic, target skill 3, fine-tuning | training steps to validation loss 0.01: 38.1% fewer steps | Improved |
| Use Skill-it! instead of target-skill-only training | Addition synthetic skill 1 · Addition synthetic, skill 1, fine-tuning | training steps to validation loss 0.01: 59% fewer steps | Improved |
| Use Skill-it! to choose training data by graph relationships | Natural Instructions test task categories · Natural Instructions, out-of-domain | evaluation loss: 11 out of 12 task categories | Improved |
| Use Skill-it! rather than uniform sampling over data sources | RedPajama 3B-parameter model · RedPajama, LM Evaluation Harness | accuracy: higher accuracy with 1B additional tokens than uniform with 3B tokens | Improved |
| Use instruction types as skills for stratified sampling | Alpaca · Alpaca, instruction-type skills | validation loss: 0.007 points over random sampling on average | No clear effect |
| Train on a balanced mixture of LEGO skills 1, 2, and 3 instead of skill 3 alone | LEGO synthetic · LEGO synthetic, skill 3 | training steps to validation loss 0.01: 35.9% fewer training steps | Improved |
| Train on a balanced mixture of addition skills 1 and 2 instead of skill 1 alone | Addition synthetic · Addition synthetic, skill 1 | training steps to validation loss 0.01: 32% fewer training steps | Improved |
| Add English QG, Spanish QA, and English QA training data to Spanish QG | Natural Instructions · Natural Instructions, Spanish question generation | validation loss: 4.1% lower validation loss; uses 25% of the Spanish QG data | Improved |
Closest papers by meaning
- Can Models Learn Skill Composition from Examples? (NeurIPS 2024 · similarity 0.89)
- Hypothesis, Verification, and Induction: Grounding Large Language Models with Self-Driven Skill Learning (AAAI 2024 · similarity 0.89)
- MASS: Mathematical Data Selection via Skill Graphs for Pretraining Large Language Models (ICML 2025 · similarity 0.89)
- Instruct-SkillMix: A Powerful Pipeline for LLM Instruction Tuning (ICLR 2025 · similarity 0.88)
- Sloth: scaling laws for LLM skills to predict multi-benchmark performance across families (NeurIPS 2025 · similarity 0.88)
- It's All Training: A Fully Synthetic Single-Stage Recipe for LLMs (NeurIPS 2026 · similarity 0.88)
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.
