Research map / Large language models / Language, safety and interpretability
Large Language Model as Attributed Training Data Generator: A Tale of Diversity and Bias
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 |
|---|---|---|---|
| replace simple class-conditional prompt with attributed prompts (AttrPrompt) | ChatGPT-generated training data, BERT-base · NYT/Amazon/Reddit/StackExchange | accuracy: NYT 75.47 -> 81.30; Amazon 57.34 -> 66.08; Reddit 53.48 -> 63.33; StackExchange 42.88 -> 48.99 | Helped |
| add class-dependent attribute value filtering (CAF) (ablation) | AttrPrompt · NYT/Amazon/Reddit/StackExchange | accuracy: NYT 80.40 -> 81.30; Amazon 61.67 -> 66.08; Reddit 61.22 -> 63.33; StackExchange 45.90 -> 48.99 | Helped |
| fix one attribute dimension instead of random attributes (ablation) | AttrPrompt · NYT etc. | accuracy | Hurt |
| use MetaPrompt instead of SimPrompt | data generation · NYT/Amazon/Reddit/StackExchange | accuracy: NYT 75.47 -> 79.58; Amazon 57.34 -> 56.35 | Mixed |
| merge generated data with Gold training set (AttrPrompt) | BERT-base · NYT/Amazon/Reddit/StackExchange | accuracy: NYT +3.67; Amazon +1.72; Reddit +1.86; StackExchange +9.30 | Helped |
| merge SimPrompt data with Gold training set | BERT-base · Amazon | accuracy: -0.38 | Hurt |
| plug AttrPrompt into SuperGen/SunGen/ProGen/ReGen | existing generation methods · NYT/Amazon | accuracy: SuperGen NYT 76.11 -> 82.05; ReGen Amazon 34.70 -> 58.40 | Helped |
| use AttrPrompt instead of SimPrompt for multi-label generation | arXiv 98 classes · arXiv | Micro F1: 26.75 -> 37.88 | Helped |
| use AttrPrompt instead of SimPrompt (diversity) | generated data · NYT | vocabulary size / INGF: INGF 5277.2 -> 6688.6 | Helped |
| AttrPrompt budget at 5% vs SimPrompt at 100% | ChatGPT querying cost · four datasets | accuracy | Helped |
| use ChatGPT instead of text-curie-001 as generator | SimPrompt · four datasets | avg accuracy | Hurt |
| AttrPrompt vs LLM zero-shot prediction | topic classification · NYT/Amazon/Reddit/StackExchange | accuracy: Reddit 63.33 vs 67.00 | Mixed |
Closest papers by meaning
- Diversity as a Reward: Fine-Tuning LLMs on a Mixture of Domain-Undetermined Data (NeurIPS 2025 · similarity 0.89)
- Fixing It in Post: A Comparative Study of LLM Post-Training Data Quality and Model Performance (NeurIPS 2025 · similarity 0.89)
- SimpleStrat: Diversifying Language Model Generation with Stratification (NeurIPS 2025 · similarity 0.88)
- Supervised Knowledge Makes Large Language Models Better In-context Learners (ICLR 2024 · similarity 0.88)
- Automating Dataset Updates Towards Reliable and Timely Evaluation of Large Language Models (NeurIPS 2024 · similarity 0.88)
- Get more for less: Principled Data Selection for Warming Up Fine-Tuning in LLMs (ICLR 2024 · 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.