Datasets:
benchmark_family stringclasses 1
value | mode stringclasses 1
value | model stringclasses 2
values | rows int64 2k 2k | seed int64 42 42 | bootstrap_replicates int64 200 200 | roc_auc float64 0.79 0.82 | pr_auc float64 0.75 0.8 | f1 float64 0.71 0.71 | log_loss float64 0.53 0.58 | brier float64 0.18 0.2 | ece float64 0.07 0.08 | threshold float64 0.4 0.5 | training_steps int64 20 49 | training_seconds float64 0.02 0.09 | inference_seconds float64 0 0 | peak_rss_bytes int64 250M 261M | estimated_communication_bytes int64 5.34M 6.99M | communication_message_count int64 597 2.36k | communication_scalar_count int64 668k 873k | forward_communication_bytes int64 2.52M 3.69M | backward_communication_bytes int64 1.65M 4.47M | traffic_type stringclasses 1
value | raw_features_pooled bool 1
class | source_type stringclasses 1
value | source_commit stringclasses 1
value |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
core_model_comparison | synthetic_scale | logistic | 2,000 | 42 | 200 | 0.822736 | 0.799092 | 0.714801 | 0.526972 | 0.175765 | 0.065591 | 0.495 | 49 | 0.016098 | 0.000113 | 249,626,624 | 5,340,000 | 597 | 667,500 | 3,693,600 | 1,646,400 | SIMULATED PAYLOAD SIZE | false | fully_synthetic | ede09d38933a91aec31296b70865a5528c4ec958 |
core_model_comparison | synthetic_scale | vfl-hist-gbdt | 2,000 | 42 | 200 | 0.788194 | 0.745349 | 0.714734 | 0.579611 | 0.196946 | 0.082897 | 0.395 | 20 | 0.092116 | 0.001982 | 261,001,216 | 6,987,376 | 2,361 | 873,422 | 2,516,600 | 4,470,776 | SIMULATED PAYLOAD SIZE | false | fully_synthetic | ede09d38933a91aec31296b70865a5528c4ec958 |
VertiMosaic VFL Benchmark
Five reproducible experiment families for exploring vertical federated learning on heterogeneous tabular data.
🚀 Open the interactive Space · 🧩 Reference models · 💻 Source code
This Hub dataset is an evaluation and reproducibility artifact, not a training corpus. It packages the canonical evidence generated by VertiMosaic's fully synthetic, CPU-first VFL benchmark so each experiment family can be inspected independently in Dataset Viewer.
At a glance
| Item | Published evidence |
|---|---|
| Core comparison | 2000 synthetic aligned entities |
| Robustness studies | 1200 entities per study |
| Experiment seed | 42 |
| Bootstrap replicates | 200 for the core comparison |
| Dataset Viewer configs | 5 |
| Raw feature pooling | No |
| Third-party source rows redistributed | No |
| Source commit | ede09d38933a91aec31296b70865a5528c4ec958 |
Pick a question
| Question | Configuration | What to inspect |
|---|---|---|
| How do the reference VFL models compare? | core-model-comparison |
ROC-AUC, PR-AUC, F1, calibration, timing and communication |
| Which parties contribute signal? | party-ablation |
Bank plus valid subsets of Telecom, Insurance and Retail |
| What changes when entity overlap is incomplete? | partial-overlap |
intersection-only vs. missing-party-aware training |
| How robust is inference to unavailable parties? | party-dropout |
training/inference availability scenarios |
| What happens under controlled feature drift? | feature-drift |
mean, variance, missingness and categorical-frequency shifts |
Dataset Viewer configurations
The card maps each configuration to exactly one canonical Parquet file so auxiliary provenance files are not inferred as splits.
core-model-comparison—data/core_model_comparison.parquet(default)party-ablation—data/party_ablation.parquetpartial-overlap—data/partial_overlap.parquetparty-dropout—data/party_dropout.parquetfeature-drift—data/feature_drift.parquet
Quick start
from datasets import load_dataset
core = load_dataset(
"sauravsingla08/VertiMosaic-VFL-Benchmark",
"core-model-comparison",
split="benchmark",
)
overlap = load_dataset(
"sauravsingla08/VertiMosaic-VFL-Benchmark",
"partial-overlap",
split="benchmark",
)
print(core.to_pandas())
print(overlap.to_pandas())
For a visual walkthrough of the same evidence, open the interactive VertiMosaic Space.
Evidence and provenance
Every published benchmark row carries the fixed experiment seed, source type, raw_features_pooled boundary and Git source commit. summaries/manifest.json records the canonical generation parameters and timestamp.
- Generated at:
2026-10-02T07:21:02.844492+00:00 - Source commit:
ede09d38933a91aec31296b70865a5528c4ec958 - Source repository: https://github.com/sauravsingla/VertiMosaic
- Reference model bundle: https://huggingface.co/sauravsingla08/VertiMosaic-VFL-Reference-Models
Timing and RSS fields are environment-dependent measurements from the GitHub Actions runner used for that publication. They should not be interpreted as hardware-independent performance claims.
Privacy boundary
VertiMosaic keeps party feature matrices inside party objects during the reference VFL protocols and records communication metadata. This research simulator does not claim cryptographic security, private set intersection, MPC, homomorphic encryption, secure aggregation, collusion resistance, malicious-party security, or formal differential privacy.
External-source policy
No UCI, OpenML, or IEEE-CIS source rows are redistributed in this Hugging Face dataset. The Apache-2.0 declaration applies to VertiMosaic-generated benchmark artifacts and does not relicense third-party data.
The separate four-industry external research benchmark is explicitly semi-synthetic: its public Bank, Telecom, Insurance and Retail sources do not represent the same real people. The optional IEEE-CIS mode uses user-supplied authorized local files and does not redistribute those files here.
Intended use
Use this dataset to reproduce and inspect VertiMosaic's VFL behavior, communication cost, party contribution, entity-overlap robustness, party availability and controlled drift experiments. It should not be treated as evidence of production privacy, fairness, security, regulatory compliance or real-world causal effects.
- Downloads last month
- 260