Research map / Large language models / AI and society
Data Portraits: Recording Foundation Model Training Data
Marc Marone, Benjamin Van Durme
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 Bloom filter instead of full text search (200 chars) | membership classifier · Pile vs NYT classification | sec/doc: 11.28 -> 0.015 | Improved |
| use Bloom filter instead of full text search (200 chars) | membership classifier · Pile vs NYT classification | F1: 1.0 -> 1.0 | No clear effect |
| shorten FTS query context from 200 to 50 characters (ablation) | FTS classifier · Pile vs NYT classification | F1: 1.0 -> 0.57 | Hurt |
| shorten FTS query context from 200 to 50 characters (ablation) | FTS classifier · Pile vs NYT classification | sec/doc: 11.28 -> 1.81 | Improved |
| use strided Bloom filter instead of suffix array | membership tool · space vs dataset | size ratio: 1.92 -> 0.03 | Improved |
| use strided Bloom filter instead of BM25 full text index | membership tool · space vs dataset | size ratio: 1.80 -> 0.03 | Improved |
| lower false positive rate from 1e-3 to 1e-4 (ablation) | Pile Bloom filter · structure size | GB: 27 -> 36 | Hurt |
| raise false positive rate from 1e-3 to 1e-2 (ablation) | Pile Bloom filter · structure size | GB: 27 -> 18 | Improved |
| shorter n-gram width (50 -> 25 characters) (ablation) | n-gram fingerprint · duplicated n-grams in Pile subset | % duplicated: 1.6 -> 7.5 | Hurt |
| shorter n-gram width (50 -> 10 characters) (ablation) | n-gram fingerprint · duplicated n-grams in Pile subset | % duplicated: 1.6 -> 79 | Hurt |
| search raw LDC AMR2.0 text instead of detokenized plaintext (ablation) | Bloom filter overlap · AMR2.0 vs Pile | %E.O.: 8.06% -> 8.12% | Improved |
Closest papers by meaning
- Position: Data Authenticity, Consent, & Provenance for AI are all broken: what will it take to fix them? (ICML 2024 · similarity 0.85)
- Navigating Dataset Documentations in AI: A Large-Scale Analysis of Dataset Cards on HuggingFace (ICLR 2024 · similarity 0.84)
- Unearthing Skill-level Insights for Understanding Trade-offs of Foundation Models (ICLR 2025 · similarity 0.84)
- CIFAR-10-Warehouse: Broad and More Realistic Testbeds in Model Generalization Analysis (ICLR 2024 · similarity 0.83)
- STAMP Your Content: Proving Dataset Membership via Watermarked Rephrasings (ICML 2025 · similarity 0.83)
- Bridging the Data Provenance Gap Across Text, Speech, and Video (ICLR 2025 · similarity 0.83)
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.
