SynBench: A Benchmark for Differentially Private Text Generation

Synthetic text generated with differential privacy (DP) from six domain-specific classification datasets, together with the original splits and the membership-inference audit (MIA) data used in the NeurIPS 2026 Evaluations & Datasets paper SynBench: A Benchmark for Differentially Private Text Generation.

Three further benchmark datasets (PsyTAR, N2C2'08, MIMIC) are under data use agreements and are not included; see the paper for how to obtain them.

Datasets and licenses

Each config is one source dataset. The original records and every synthetic dataset derived from them are released under the license of that source dataset — the licenses differ, so check the one for the config you use. The repository-level license: other reflects this.

config	name in paper	domain	task	train / test	source	license
hallmarks_of_cancer	HoC	biomedical	multi-label	10301 / 1818	bigbio/hallmarks_of_cancer	gpl-3.0
DMSAFN	DMSAFN	finance	single-label (3 classes)	3876 / 970	Daniel-ML/sentiment-analysis-for-financial-news-v2	cc-by-nc-sa-4.0
Lc37Fnd	LC37FND	finance	single-label (3 classes)	5620 / 1424	luckycat37/financial-news-dataset	cc-by-nc-sa-4.0
ZSTFNS	ZSTFNS	finance	single-label (3 classes)	9543 / 2388	zeroshot/twitter-financial-news-sentiment	mit
AYSLAX	ASYLAX	legal	single-label (3 classes)	7999 / 1000	clairebarale/AsyLex	cc-by-nc-sa-4.0
EURLEX	EURLEX	legal	multi-label (8 EuroVoc concepts)	4942 / 1245	EURLEX57K	cc-by-sa-4.0

Loading

from datasets import load_dataset

ds = load_dataset("Ydsun/SynBench", "DMSAFN")
ds["train"]                                   # real training split
ds["test"]                                    # real test split (held out, never used for generation)
ds["llama_3_2_1b__dp_transformers__eps_1"]    # synthetic data, DP-Gen on Llama-3.2-1B, ε = 1

Every record is {"text": str, "label": list[str]}; single-label datasets use a one-element list.

Split names

Synthetic splits are named <generator>__<method>__eps_<ε> (lower case, non-alphanumerics → _):

• generator: llama_3_2_1b (main results), llama_3_2_1b_instruct, llama_3_2_3b_instruct, llama_3_1_8b_instruct (generator-size study; AYSLAX and DMSAFN only), olmo_2_0425_1b_10epochs (OLMo-2 replication, DP-Gen only, generator fine-tuned for 10 epochs)
• method: dp_transformers (DP-Gen), aug_pe (AUG-PE), dp_diffusion (DP-Dfs), epsvec (EPSVec)
• ε: 0_5, 1, 2, 4, or inf (no DP). In the raw file names no-DP is nodp for Llama and 0 for OLMo-2.

Repository layout

<dataset>/
├── original/                      train-original.jsonl, test-original.jsonl
├── Llama-3.2-1B/
│   ├── dp-transformers/           train{0.5,1,2,4,nodp}-dp-transformers.jsonl
│   ├── aug-pe/                    train{…}-aug-pe.jsonl
│   ├── dp-diffusion/              train{…}-dp-diffusion.jsonl
│   └── EPSVec/                    train{…}-EPSVec.jsonl
├── Llama-3.2-{1B,3B}-Instruct/, Llama-3.1-8B-Instruct/   (AYSLAX, DMSAFN)
│   └── dp-transformers/           train{…}-dp-transformers-<model>.jsonl
├── OLMo-2-0425-1B-10epochs/
│   └── dp-transformers/           train{0,0.5,1,2,4}-dp-transformers.jsonl   (0 = no DP)
└── mia/                           (AYSLAX, DMSAFN)
    ├── vulsamples/                all_outliers.jsonl (MIA targets), outlier_indices.npy, config.json
    ├── D_sample/
    │   ├── split_datasets/        D_p_indices.npy, D_a_indices.npy (private / auxiliary split of train)
    │   ├── sampled_private_datasets/   dataset_{0..99}.jsonl + sampling_labels.json (membership ground truth)
    │   └── sampled_4ref_datasets_outliers/  dataset_{0..3}.jsonl + sampling_labels.json (reference sets)
    └── D_synth/{augpe,dp-transformers}/finalresults/
                                   dataset_{i}_e{ε}_samples.jsonl  (synthetic version of private subset i; e0 = no DP)

Synthetic data: record counts

Most generators produce one record per real training record. Counts differ where generation or post-processing dropped records (OLMo-2 DP-Gen outputs were de-duplicated on (text, label), keeping the first occurrence; AUG-PE and DP-Dfs drop empty or invalid generations). "–" = not generated.

hallmarks_of_cancer

generator	method	ε=inf	ε=4	ε=2	ε=1	ε=0.5
Llama-3.2-1B	dp-transformers	10301	10301	10301	10301	10301
Llama-3.2-1B	aug-pe	9550	9451	9497	9471	9410
Llama-3.2-1B	dp-diffusion	10300	10299	10273	10067	9660
Llama-3.2-1B	EPSVec	10301	10301	10301	10301	10301
OLMo-2-0425-1B-10epochs	dp-transformers	8287	10263	10262	10226	10174

DMSAFN

generator	method	ε=inf	ε=4	ε=2	ε=1	ε=0.5
Llama-3.2-1B	dp-transformers	3876	3876	3876	3876	3876
Llama-3.2-1B	aug-pe	3687	3681	3677	3693	3684
Llama-3.2-1B	dp-diffusion	3876	3876	3876	3638	24
Llama-3.2-1B	EPSVec	3876	3876	3876	3876	3876
Llama-3.2-1B-Instruct	dp-transformers	3872	3876	3876	3876	3876
Llama-3.2-3B-Instruct	dp-transformers	3857	3876	3876	3876	3876
Llama-3.1-8B-Instruct	dp-transformers	3772	3876	3876	3876	3876
OLMo-2-0425-1B-10epochs	dp-transformers	2875	3825	3839	3832	3762

Lc37Fnd

generator	method	ε=inf	ε=4	ε=2	ε=1	ε=0.5
Llama-3.2-1B	dp-transformers	5349	5609	5615	5616	5597
Llama-3.2-1B	aug-pe	5614	5614	5614	5614	5614
Llama-3.2-1B	dp-diffusion	5620	5620	5620	5620	5620
Llama-3.2-1B	EPSVec	5620	5620	5620	5620	5620
OLMo-2-0425-1B-10epochs	dp-transformers	5620	5608	5617	5606	5608

ZSTFNS

generator	method	ε=inf	ε=4	ε=2	ε=1	ε=0.5
Llama-3.2-1B	dp-transformers	6710	9435	9370	9001	9541
Llama-3.2-1B	aug-pe	9010	8911	8850	8836	8778
Llama-3.2-1B	dp-diffusion	9537	9543	9543	9437	899
Llama-3.2-1B	EPSVec	9543	9543	9543	9543	9543
OLMo-2-0425-1B-10epochs	dp-transformers	7353	9535	9522	9525	9511

AYSLAX

generator	method	ε=inf	ε=4	ε=2	ε=1	ε=0.5
Llama-3.2-1B	dp-transformers	8999	8994	8995	8989	8998
Llama-3.2-1B	aug-pe	7958	7734	7730	7713	7746
Llama-3.2-1B	dp-diffusion	7999	7999	7999	7999	7999
Llama-3.2-1B	EPSVec	7999	7999	7999	7999	7999
Llama-3.2-1B-Instruct	dp-transformers	7999	7999	7999	7999	7999
Llama-3.2-3B-Instruct	dp-transformers	7999	7999	7999	7999	7999
Llama-3.1-8B-Instruct	dp-transformers	7999	7999	7996	7999	7999
OLMo-2-0425-1B-10epochs	dp-transformers	7999	7997	7994	7995	7995

EURLEX

generator	method	ε=inf	ε=4	ε=2	ε=1	ε=0.5
Llama-3.2-1B	dp-transformers	4942	4942	4942	4942	4942
Llama-3.2-1B	aug-pe	4858	4857	4858	4856	4858
Llama-3.2-1B	dp-diffusion	4942	4561	4082	2768	380
Llama-3.2-1B	EPSVec	4942	4942	4942	4942	4942
OLMo-2-0425-1B-10epochs	dp-transformers	4933	4939	4938	4938	4934

How the data was generated

• DP-Gen (dp-transformers): the generator is fine-tuned with DP-SGD conditioned on the label as a control code, then sampled with one prompt per real training record's label set.
• AUG-PE (aug-pe): private evolution with the generator as the API model.
• DP-Dfs (dp-diffusion): DP-SGD-trained text diffusion with a GENIE backbone (independent of the Llama/OLMo generator; stored under Llama-3.2-1B/ for layout only).
• EPSVec (EPSVec): steering with a privatised dataset vector.

ε ∈ {0.5, 1, 2, 4} plus a no-DP run. Full hyper-parameters are in the paper's appendix A.3.

Dataset-specific notes

• EURLEX is a subset of EURLEX57K restricted to 8 EuroVoc concepts (see the paper for the selection).
• LC37FND labels are derived from the sentiment scores of luckycat37/financial-news-dataset (articles scraped from finance.yahoo.com); see the paper for the mapping.
• ASYLAX original texts contain mis-decoded UTF-8 (e.g. ’); they are released unchanged because all generators and classifiers were run on them.
• DP-Dfs collapses at small ε on some datasets (e.g. DMSAFN ε=0.5: 24 records; EURLEX ε=0.5: 380; ZSTFNS ε=0.5: 899). These files are released as produced; the paper reports such settings as unusable.
• AYSLAX Llama-3.2-1B/dp-transformers files have ~8,990–8,999 records (more than the 7,999 training records); they are released as produced.
• AYSLAX mia/D_synth/augpe/finalresults lacks dataset_1_e0.5 and dataset_1_e2.
• mia/D_synth/augpe files originally stored label as a string; in this release it is wrapped in a one-element list to match every other file. Texts are unchanged.
• OLMo-2 10-epoch DP-Gen files had an extra prompt column equal to the generation control code ("\t".join(label) + "\n\n"); it is dropped here so all splits share the text/label schema.

Intended use

Benchmarking DP synthetic text generators (utility: train a classifier on a synthetic split, evaluate on test; fidelity: compare synthetic to real text; privacy: re-run the MIA audit). The synthetic data is not a privacy-safe substitute for the originals here — the originals are public and are released alongside it.

Citation

@inproceedings{sun2026synbench,
  title     = {SynBench: A Benchmark for Differentially Private Text Generation},
  author    = {Sun, Yidan and Schlegel, Viktor and Nandakumar, Srinivasan and Zahid, Iqra and Wu, Yuping and
               Wu, Yulong and Li, Hao and Zhang, Jie and Del-Pinto, Warren and Nenadic, Goran and
               Lam, Siew Kei and Bharath, Anil Anthony},
  booktitle = {Advances in Neural Information Processing Systems (Evaluations and Datasets Track)},
  year      = {2026}
}

Please also cite the source dataset of every config you use (see the table above).