Research map / Large language models / Retrieval and knowledge
A Fictional Q&A Dataset for Studying Memorization and Knowledge Acquisition
John Kirchenbauer, Natjanan Mongkolsupawan, Yuxin Wen, Tom Goldstein, Daphne Ippolito
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 |
|---|---|---|---|
| fine-tune on fictional documents (5% of each batch) | Llama-3.1-8B · FictionalQA Event Split MCQ, train-event questions | MCQ accuracy: 31.91% -> 45.31% | Improved |
| fine-tune on fictional documents | Llama-3.1-8B · FictionalQA Event Split MCQ, held-out-event questions | MCQ accuracy: 30.51% -> 43.44% | Improved |
| train on Doc Split instead of Event Split (ablation) | LLaMA, Gemma models · FictionalQA validation loss | validation loss minimum | Improved |
| train on structured Fictsheets instead of diverse documents (ablation) | Llama-3.1-8B · FictionalQA MCQ | MCQ accuracy | Hurt |
| train on Fictsheets split (ablation) | LLaMA, Gemma models · FictionalQA validation loss | validation loss | Hurt |
| increase model size (ablation) | Llama and Gemma models · FictionalQA MCQ | MCQ accuracy | Improved |
| train on fictional documents | Llama-3.1-8B · TriviaQA answer NLL | answer NLL | No clear effect |
| use Style split (hold out news or blog) instead of Doc Split (ablation) | Llama-3.1-8B · FictionalQA Q&A transfer | Q&A transfer | No clear effect |
| raise fiction sampling rate to 50% or 100% (ablation) | fine-tuning mixture · FictionalQA | generalization | Hurt |
Closest papers by meaning
- How Do Large Language Models Acquire Factual Knowledge During Pretraining? (NeurIPS 2024 · similarity 0.89)
- Generalization v.s. Memorization: Tracing Language Models’ Capabilities Back to Pretraining Data (ICLR 2025 · similarity 0.88)
- Physics of Language Models: Part 3.1, Knowledge Storage and Extraction (ICML 2024 · similarity 0.87)
- Co-occurrence is not Factual Association in Language Models (NeurIPS 2024 · similarity 0.87)
- Long-form factuality in large language models (NeurIPS 2024 · similarity 0.87)
- Copyright Traps for Large Language Models (ICML 2024 · similarity 0.86)
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.
