Research map / Large language models / Language, safety and interpretability
SILO Language Models: Isolating Legal Risk In a Nonparametric Datastore
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Trains 1.3B LMs only on permissively licensed text (Open License Corpus, 228B tokens) and adds an inference-time nonparametric datastore (kNN-LM or RIC-LM) holding high-risk text, closing ~90% of the perplexity gap to Pythia-1.4B.
Problem
Language models trained only on low-risk (public domain / permissively licensed) text degrade substantially out-of-domain, creating a tradeoff between copyright/legal risk and model performance.
Approach
Split training data: learn parametric LM weights only on permissive data (OLC), and place high-risk data in a nonparametric datastore queried at inference via kNN-LM or RIC-LM, enabling attribution and per-example opt-out.
Key results
- Parametric-only SILO (PDSW) averages 17.3 perplexity vs 11.4 for Pythia-1.4B across 14 domains; out-of-domain gaps are large (e.g., Amazon 34.8 vs 22.9).
- kNN-LM datastore reduces the average gap to Pythia by 91% on 8 out-of-domain datasets (SILO PDSW kNN-LM 12.0 vs Pythia 11.4 ppl).
- RIC-LM reduces the gap by only 27% on average; kNN-LM scales better with datastore size in a log-linear trend.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add a kNN-LM datastore | SILO PD+SW · Average (8 out-of-domain datasets) | perplexity: 19.0 -> 12.0 | Helped |
| Add a RIC-LM datastore | SILO PD+SW · Average (8 out-of-domain datasets) | perplexity: 19.0 -> 16.9 | Helped |
| Use kNN-LM rather than RIC-LM retrieval | SILO PD+SW · Average (8 out-of-domain datasets) | perplexity: 16.9 -> 12.0 | Helped |
| Add source code to PD+SW training data (ablation) | SILO parametric LM · PhilPapers validation | perplexity: 36.4 -> 24.2 | Helped |
| Upsample low-resource domains during training (ablation) | SILO PD+SW parametric LM · CC News validation | perplexity: 21.8 -> 21.3 | Helped |
| Upsample low-resource domains during training (ablation) | SILO PD+SW parametric LM · FreeLaw validation | perplexity: 4.9 -> 5.7 | Hurt |
| Use concat-2 instead of basic RIC-LM (ablation) | SILO PD+SW · CC News validation | perplexity: 14.8 -> 17.0 | Hurt |
| Use concat-next instead of basic RIC-LM (ablation) | SILO PD+SW · CC News validation | perplexity: 14.8 -> 18.8 | Hurt |
| Ensemble predictions from the top 10 retrieved blocks instead of using the top block (ablation) | SILO PD+SW RIC-LM · CC News validation | perplexity: 14.8 -> 13.5 | Helped |
| Use quantized-vector distance approximation instead of original embeddings (ablation) | SILO kNN-LM · Language modeling | perplexity: about 0.5% | Hurt |
Surprising findings
- Most remaining degradation comes from the out-of-domain parametric LM rather than the out-of-domain kNN encoder, which is robust to domain shift.
- Parametric LMs trained on PD only achieve high perplexity on HackerNews and Github despite strong in-domain performance elsewhere, showing extreme domain generalization difficulty.
Limitations
- OLC relies on explicit license metadata, so the permissive corpus may be underestimated and not fully safe; the paper states it does not eliminate legal risk.
- Remaining gap to Pythia persists on some domains (e.g., Books3, NIH ExPorter, Wikipedia) and the datastore is capped at about 1B tokens.
- Potential fairness issues from copyright-expired older books in training data.
Experimental setup
Datasets: Open License Corpus (OLC) The Pile validation/test domains IMDB (downstream datastore) Amazon reviews MIMIC-III Enron Emails
Benchmarks: The Pile domains (14 perplexity test sets) AGNews Yahoo Subj SST-2 MR RT CR Amazon polarity RTE
Metrics: perplexity zero-shot accuracy
Baselines: Pythia-1.4B RIC-LM kNN-LM parametric-only SILO
Builds on: kNN-LM retrieval-in-context LM LLaMA OpenLM
Closest papers by meaning
- The Common Pile v0.1: An 8TB Dataset of Public Domain and Openly Licensed Text (NeurIPS 2025 · similarity 0.87)
- SaulLM-54B & SaulLM-141B: Scaling Up Domain Adaptation for the Legal Domain (NeurIPS 2024 · similarity 0.87)
- Evaluating Copyright Takedown Methods for Language Models (NeurIPS 2024 · similarity 0.86)
- DOMBA: Double Model Balancing for Access-Controlled Language Models via Minimum-Bounded Aggregation (AAAI 2025 · similarity 0.86)
- SafetyPrompts: A Systematic Review of Open Datasets for Evaluating and Improving Large Language Model Safety (AAAI 2025 · similarity 0.86)
- Combating Data Laundering in LLM Training (NeurIPS 2026 · 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.