Research map / Large language models / Instruction and fine-tuning

HARMONIC: Harnessing LLMs for Tabular Data Synthesis and Privacy Protection

NeurIPS 2024 · Large language models › Instruction and fine-tuning · approach: Fine-tuning generalization and delta compression · 5 citations (OpenAlex)

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 HARMONIC, which instruction-fine-tunes LLaMA-2-7b-chat with kNN-grouped table rows to generate privacy-preserving synthetic tabular data, and introduces the DLT and LLE evaluation metrics.

Problem

Generating realistic synthetic tabular data with LLMs while limiting privacy leakage of the training rows, and evaluating synthetic data and privacy risks in LLM-based downstream scenarios.

Approach

Instead of continued pre-training on serialized rows, fine-tune a larger LLM on instruction examples built from kNN groups of k+1 rows (k=5), computing loss only on the output row and shuffling feature order; evaluate with LLE (downstream LLM fine-tuning efficacy) and DLT (perplexity gap between train and synthetic data).

Key results

• On LLE, HARMONIC scores 0.64 on GM vs 0.71 for real data and 0.75 on DI vs 0.70 for real data, with the authors noting an average drop under 5% versus real data.
• TABSYN achieves the best LLE on GM (0.72) and DI (0.77), and the authors state it beats real data only on GM and DI.
• On privacy, HARMONIC has the highest DCR on GM (8.08) and DLT of -0.16 on GM, and the authors claim it surpasses or is second on most datasets.

Measured effects reported in the paper

Surprising findings

• Pretraining-style GReaT and REaLTabFormer show very negative DLT values (e.g., -163.71 for RTF on AD), which the authors interpret as significant privacy risk.
• Their own MLE metric results show HARMONIC lower than real data on several datasets, while LLE shows smaller gaps, suggesting traditional MLE evaluation may be misleading for LLM use.

Limitations

• Longer processing and sampling time than smaller-model baselines.
• Limited to classification tasks; LLMs are less suited to numerical regression.
• Privacy claims rest on a proposed perplexity-based DLT metric and qualitative kNN-structure arguments without formal privacy guarantees.

Experimental setup

Datasets: German Credit Adult Income Diabetes Buddy

Benchmarks: German Credit Adult Income Diabetes Buddy

Metrics: LLE weighted F1 MLE F1 NRS DCR DLT

Baselines: SMOTE TVAE CTAB-GAN TabDDPM TABSYN GReaT REaLTabFormer

Builds on: GReaT TABSYN LoRA k-nearest neighbors

Closest papers by meaning

• Breaking the Quality–Privacy Tradeoff in Tabular Data Generation via In-Context Learning (NeurIPS 2026 · similarity 0.89)
• Do You Really Need Public Data? Surrogate Public Data for Differential Privacy on Tabular Data (NeurIPS 2025 · similarity 0.87)
• CuTS: Customizable Tabular Synthetic Data Generation (ICML 2024 · similarity 0.86)
• Open LLMs are Necessary for Current Private Adaptations and Outperform their Closed Alternatives (NeurIPS 2024 · similarity 0.86)
• The Canary’s Echo: Auditing Privacy Risks of LLM-Generated Synthetic Text (ICML 2025 · similarity 0.86)
• Privacy-Preserving Data Release Leveraging Optimal Transport and Particle Gradient Descent (ICML 2024 · similarity 0.86)