Research map / Large language models / Instruction and fine-tuning
Data Provenance Auditing of Fine-Tuned Large Language Models with a Text-Preserving Technique
Yanming Li, Cédric Eichler, Nicolas Anciaux, Alexandra Bensamoun, Lorena Gonzalez-Manzano, Seifeddine Ghozzi
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 |
|---|---|---|---|
| increase number of watermarked documents (ablation) | invisible Unicode cue-reply watermark · Blog1k, Poems, News | regurgitation rate: Mistral Blog1k p 0.060 -> 0.431 (30 -> 50 docs) | Improved |
| increase amount of unmarked fine-tuning documents (ablation) | invisible Unicode watermark · Blog1k, Poems, News | regurgitation rate: Mistral News 0.74 at 3,000 docs vs about 0.25 at 1,000 and 9,000 | No clear effect |
| increase number of distinct watermarks in fine-tuning data (1 to 50) (ablation) | invisible Unicode watermark · Blog1k, Poems, News | regurgitation rate: Mistral about 0.5-0.7 at U=1 -> above 0.85 at U=50 | Improved |
| use cue-reply watermark with ranking test against counterfactuals | Mistral-7B, LLaMA-2-7B fine-tuning audit · Blog1k, Poems, News | TPR at 0% FPR: 96.7% TPR at 0% FPR with k=1 | Improved |
| replace homoglyph watermark baseline with invisible cue-reply watermark under 50 unique watermarks | homoglyph watermarking (Wei et al., 2024) · Blog1k, Poems, News, 50 watermarks | detection significance: baseline Z-scores about -0.1 to -0.6 (p>0.2) vs regurgitation rates >0.8 | Improved |
| fine-tune GPT-OSS-20B instead of 7B models (ablation) | multi-watermark setting · Appendix D | regurgitation rate | Improved |
Closest papers by meaning
- On the Reliability of Watermarks for Large Language Models (ICLR 2024 · similarity 0.88)
- Break the Trade-off Between Watermark Strength and Speculative Sampling Efficiency for Language Models (ICLR 2026 · similarity 0.88)
- TextSeal: A Localized LLM Watermark for Provenance & Distillation Protection (NeurIPS 2026 · similarity 0.87)
- Can Watermarking Large Language Models Prevent Copyrighted Text Generation and Hide Training Data? (AAAI 2025 · similarity 0.87)
- Watermarking Makes Language Models Radioactive (NeurIPS 2024 · similarity 0.87)
- On the Learnability of Watermarks for Language Models (ICLR 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.
