zlysunshine/aave-bns-data

Aave-BNS Multidimensional Protocol-Network Evidence

Hugging Face release status: PRIVATE RELEASE CANDIDATE. The dataset has been uploaded to the Hugging Face Hub for post-upload validation. Local validation covers 14 configurations, Croissant 1.1 core metadata, substantive Responsible AI metadata, checksums, and deterministic handoff auditing. Dataset Viewer validation, clean Hub loading, platform-generated Croissant reconciliation, and immutable release tagging are the remaining publication gates.

Evaluative role

This resource supports evaluation of how decentralization claims change when participation, Pool-event-frequency concentration, role-aware network structure, actor uncertainty, and infrastructure evidence are kept distinct. It is a data-centric empirical and measurement-audit resource, not a benchmark leaderboard and not a person-level dataset.

Dataset structure

Layer	Path	Contents	Release state
Queried-data evidence	data/queried/	Retrieval windows, provider comparisons, manifests, and artifact receipts recovered from completed acquisition	Included as audit evidence; no new query was run
Processed data	data/processed/	Fourteen governed Hub configurations and two audit-only evidence-gap tables	Included; aggregate or privacy-minimized
Metadata	metadata/	Schema, dictionary, sources, Croissant, RAI, quality, claims, checksums, migration ledger, and Hub handoff	Included and locally validated

The release does not expose raw address-level event payloads or natural-person labels. The migration ledger records their scientific lineage and exclusion; missing data must never be silently replaced by a fresh query.

Evidence regimes

• OBSERVED: audited protocol/source evidence.
• DERIVED: deterministic metrics computed from observed evidence.
• SYNTHETIC: uncalibrated mechanism simulation.
• BOUNDED: partial-identification outputs.
• FAILED_DESIGN: retained but rejected identification diagnostics.
• BLOCKED/PLANNED: metadata-only evidence gaps. These tables are excluded from the Hub configuration list and Croissant record sets.

Essential interpretation boundary

Address ≠ natural person ≠ user ≠ independent economic actor. Event-frequency HHI is not capital, liquidity, wealth, risk, welfare, ownership, or governance power. The Arbitrum–Gnosis comparison is a failed-donor arithmetic diagnostic and is not a treatment effect. Verified route-level infrastructure dependence is not measured.

Configuration inventory

Each configuration contains one complete governed table exposed under the Hub packaging split name train; this label does not imply an ML training partition. Primary keys, types, units, missing-value semantics, provenance, and claim mappings are machine-readable in metadata/release_manifest.json, metadata/schema.json, metadata/data_dictionary.csv, and metadata/result_data_crosswalk.csv.

Configuration	Rows	Evidence regime	Observational unit
observed_protocol_action_counts	165	DERIVED	chain-week-action cohort
treatment_event_registry	2	OBSERVED	governed protocol event
participation_and_concentration_metrics	66	DERIVED	chain-week
cross_chain_overlap	33	DERIVED	week
structural_metrics	12	DERIVED	chain-layer
actor_bounds_period	4	BOUNDED	analysis period
actor_bounds_change	4	BOUNDED	comparison-metric-assumption
simulation	16	SYNTHETIC	scenario-parameter setting
failed_design_estimates	6	FAILED_DESIGN	outcome-horizon diagnostic
failed_design_event_study	66	FAILED_DESIGN	outcome-event-week diagnostic
failed_design_placebos	6	FAILED_DESIGN	outcome-pseudo-event diagnostic
failed_design_pretrends	2	FAILED_DESIGN	outcome pretrend diagnostic
source_audit	31	OBSERVED	evidence event
metric_registry	39	OBSERVED	governed measurement definition

Loading and local validation

Install the validation dependencies and run the fail-closed handoff checks:

python3.11 -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements-validation.txt
python scripts/validate_release.py . --json-out metadata/validation_run.json
python scripts/clean_load.py . --require-hf-datasets
python scripts/validate_hf_handoff.py .

Load a local configuration:

from datasets import load_dataset

counts = load_dataset(
    ".",
    "observed_protocol_action_counts",
    split="train",
)

After upload, replace . with zlysunshine/aave-bns and pin the immutable Hub revision recorded in the publication receipt.

Hub upload handoff

The repository README.md is the Dataset Card that the Hub will render. The declared CSV configurations allow the Dataset Viewer to infer schemas and, for Hub-supported conversion, generate a platform Croissant core. Upload only from a clean, validated checkout:

huggingface-cli upload zlysunshine/aave-bns . . \
  --repo-type dataset \
  --commit-message "Stage Aave-BNS dataset 1.0.0-rc1"

Immediately after upload:

1. record the immutable Hub commit;
2. load all 14 configurations from that revision in a clean environment;
3. inspect every Viewer configuration;
4. retrieve the Hub-generated Croissant core;
5. merge the verified RAI and PROV fields from metadata/croissant.json;
6. validate the merged file using the NeurIPS-linked validator; and
7. write the results to a publication receipt without overwriting this staged audit history.

Responsible use

Supported uses include reproducing descriptive and partial-identification results, auditing multidimensional DeFi-network measures, checking metric implementations, and inspecting failed-design diagnostics. Unsupported uses include re-identifying address holders, inferring personal traits, treating addresses as people, ranking chains with one scalar decentralization score, presenting diagnostic coefficients as causal effects, or inferring wealth, governance power, or infrastructure resilience.

The released tables contain aggregates, metrics, registries, and source URLs. Some URLs resolve to public blockchain records containing pseudonymous addresses, so linkage and profiling risks remain even though row-level address payloads are excluded.

Biases and limitations

Coverage is protocol-, event-, chain-, and time-window-specific. Equal event weighting emphasizes repeated calls, including automated contract activity, and does not weight economic value. Chain/source availability can create temporal and infrastructure measurement skew. The synthetic configuration is an uncalibrated mechanism illustration. Causal and route-level infrastructure claims are expressly unsupported.

License and third-party rights

The original curation, metadata, and released derivatives are licensed under the Creative Commons Attribution 4.0 International License (CC-BY-4.0). Attribute the named creators, cite this dataset, link the license, and indicate changes.

This license applies only to rights the licensors are authorized to grant. It does not relicense third-party websites, protocol documentation, software, trademarks, or public-chain source records linked by the metadata; those retain their own terms and legal status.

Citation, maintenance, and versioning

Citation metadata is in CITATION.cff. Version 1.0.0-rc1 is the validated pre-upload candidate. Material releases will use semantic versions, immutable Hub commits, checksums, dated quality reports, and explicit supersession records. Security, privacy, licensing, or scientific corrections should be reported through the public data repository issue tracker.

The completed acquisition source snapshot is pinned at sunshineluyao/aave-bns@932f6f4f62c3402adf38231ed83ea9ca17cc227c.