Research map / Federated learning and privacy / Communication-efficient FL
PrE-Text: Training Language Models on Private Federated Data in the Age of LLMs
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.
Proposes PrE-Text, which generates DP synthetic text via adapted Private Evolution (RoBERTa mask-filling variations) plus LLaMA-2-7B seed expansion, and shows training on that synthetic data beats DP on-device training (DP-FedAvg/DP-FTRL) for small models and improves LLaMA-2-7B on private data.
Problem
Training models on private federated user text is usually done by on-device DP federated learning, which is communication-, computation-heavy, cannot handle LLMs, and is hard to debug and deploy.
Approach
Adapt Private Evolution to text by using masked-language-model (RoBERTa) variations, reuse synthetic samples from all PE iterations, and post-process DP seeds with an LLM (LLaMA-2-7B) to expand them into a large synthetic training set, so the server can train models (including LLMs) without on-device training.
Key results
- At epsilon=1.29, PrE-Text reaches 0.718 next-token accuracy on Jobs versus 0.703 for DP-FTRL and 0.701 for DP-FedAvg (DistilGPT2).
- PrE-Text uses about 100x less per-round communication and about 6x less client computation per round, with 9x fewer rounds than on-device DP-FL.
- LLaMA-2-7B finetuned on PrE-Text synthetic data improves private-data accuracy over the non-finetuned model on all four datasets (e.g., Jobs 0.458 to 0.523 at epsilon=7.58).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train DistilGPT2 on PrE-Text synthetic data instead of using DP-FTRL | DistilGPT2 · Jobs, Forums, Microblog, Code (ε=1.29 and ε=7.58) | accuracy: Jobs: 0.703 ± 0.000 -> 0.718 ± 0.000 (ε=1.29); 0.704 ± 0.000 -> 0.721 ± 0.000 (ε=7.58) | Helped |
| Finetune LLaMA-2-7B on PrE-Text DP synthetic data | LLaMA-2-7B · Jobs, Forums, Microblog, Code | accuracy: Jobs: 0.458 ± 0.000 -> 0.522 ± 0.000; Forums: 0.466 ± 0.000 -> 0.479 ± 0.001; Microblog: 0.473 ± 0.000 -> 0.484 ± 0.000; | Helped |
| Finetune LLaMA-2-7B on PrE-Text DP synthetic data | LLaMA-2-7B · Jobs, Forums, Microblog, Code | cross-entropy loss: Jobs: 3.252± 0.000 -> 2.811 ± 0.000; Forums: 3.250 ± 0.000 -> 3.177 ± 0.004; Microblog: 3.220 ± 0.000 -> 3.133 ± 0.002; | Helped |
| Increase synthetic expansion data size (ablation) | PrE-Text with DistilGPT2 · Jobs, Forums, Microblog, Code (ε=7.58) | accuracy | Mixed |
| Use more communication rounds for DP-FedAvg or DP-FTRL (ablation) | On-device baselines · DP-FedAvg and DP-FTRL evaluation | evaluation metric | Helped |
Surprising findings
- On the Code dataset, downstream performance shows diminishing returns around 1M synthetic samples and even decreases at 2M.
- DP-Prompt performed worse than the epsilon=0 baseline when used to finetune LLaMA-2-7B.
Limitations
- Privacy relies on pretrained models trained on public data that may be contaminated or not intended for such use; no guarantees for that public data.
- Evaluation is limited to next-token prediction loss/accuracy on four datasets, with DistilGPT2 and LLaMA-2-7B as models.
- Server-side compute is substantially higher, and LLaMA-2-7B experiments used only 50,000 samples due to compute limits.
Experimental setup
Datasets: c4-en Jobs Forums Microblog Code
Benchmarks: Jobs Forums Microblog Code
Metrics: next-token prediction accuracy cross-entropy loss
Baselines: DP-FedAvg DP-FTRL DP-Prompt c4-only Expand-only Expand-private
Builds on: Private Evolution DP-FedAvg DP-FTRL Opacus
Closest papers by meaning
- SynBench: A Benchmark for Differentially Private Text Generation (NeurIPS 2026 · similarity 0.90)
- Differentially Private Synthetic Data via Foundation Model APIs 2: Text (ICML 2024 · similarity 0.89)
- Synthesizing Privacy-Preserving Text Data via Finetuning *without* Finetuning Billion-Scale LLMs (ICML 2025 · similarity 0.89)
- Open LLMs are Necessary for Current Private Adaptations and Outperform their Closed Alternatives (NeurIPS 2024 · similarity 0.89)
- Privacy-Preserving Instructions for Aligning Large Language Models (ICML 2024 · similarity 0.88)
- Scaling Laws for Differentially Private Language Models (ICML 2025 · 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.