nicolas-brieuc/dataset-trust-auditor-events

Dataset Trust Auditor — Audit Events

Public audit trail produced by the Dataset Trust Auditor — a two-phase AI pipeline that scores HuggingFace datasets across 8 trust dimensions.

Every audit run appends one row. The dataset grows over time as users audit datasets through the deployed app.

Dataset Structure

Each row is one completed audit of a HuggingFace dataset.

Column	Type	Description
audit_id	string	UUID for this audit run
url	string	Full HuggingFace dataset URL audited
dataset_name	string	Dataset identifier in owner/name format
author	string	Dataset author/organisation from HF API
timestamp	string	ISO 8601 UTC timestamp of the audit
route	string	done (Phase 1 only) or phase2 (deep analysis triggered)
confidence	float	Routing confidence score (0–1)
preliminary_score	float	Phase 1 composite trust score (0–1)
final_score	float	Final trust score after any Phase 2 revision (0–1)
score_delta	float	Difference between final and preliminary score
dim_provenance	float	Provenance dimension score (0–1)
dim_documentation	float	Documentation dimension score (0–1)
dim_licensing	float	Licensing dimension score (0–1)
dim_community_trust	float	Community trust dimension score (0–1)
dim_freshness	float	Freshness dimension score (0–1)
dim_data_quality	float	Data quality dimension score (0–1)
dim_label_quality	float	Label quality dimension score (0–1)
dim_bias_fairness	float	Bias & fairness dimension score (0–1)
license	string	SPDX license identifier from HF API
modality	string	Detected modality: text, image, audio, tabular, or null
has_dataset_card	bool	Whether a dataset card was present
card_length_chars	int	Character count of the dataset card
downloads	int	Download count at time of audit
likes	int	Like count at time of audit
last_modified	string	ISO 8601 timestamp of last dataset modification
dataset_type	string	Detected type: classification, pretraining_corpus, qa, structured, unknown
bias_strategy	string	Bias analysis strategy applied in Phase 2
discrepancy_count	int	Number of discrepancies found between Phase 1 and Phase 2
dimensions_flagged	int	Number of dimensions with score adjustment ≥ 0.10
discrepancies	string	JSON array of discrepancy description strings
key_findings	string	JSON array of finding strings (green/amber/red signals)
score_breakdown_json	string	JSON object with all 8 dimension scores
record_type	string	Event type: audit, cache_hit, error, ask_ai
duration_ms	int	Pipeline wall-clock time in milliseconds
cache_hit	bool	Whether this result was served from cache

Scoring Pipeline

Phase 1 — Fast signal assessment evaluates all 8 dimensions from HuggingFace API metadata and community signals (downloads, likes). Takes ~5–10 seconds. Datasets scoring ≥ 0.70 receive an immediate report (route = done).

Phase 2 — Agentic deep exploration is triggered for lower-scoring datasets. Samples up to 1,000 rows of actual Parquet data and runs schema inspection, label distribution analysis, data quality checks, content quality analysis, dataset type classification, and adaptive bias analysis. Produces score adjustments and discrepancy findings. Takes 1–4 minutes.

Source

Audit events are written by two sources:

• The deployed app at dataset-trust-auditor.hub.zerve.cloud — public, no login required
• The Zerve canvas pipeline (authoritative Zerve-native environment)

Built for the Zerve Hackathon 2026 as a solo project. The pipeline, scoring rules, and governance documentation (with ADRs and issue logs) are maintained at github.com/nicolas-brieuc/dataset-trust-auditor.

This dataset is itself an example of what Dataset Trust Auditor produces — every row is a trust assessment of another HuggingFace dataset.

How to Use

from datasets import load_dataset

ds = load_dataset("nicolas-brieuc/dataset-trust-auditor-events", split="train")
df = ds.to_pandas()

# Datasets that triggered Phase 2
phase2 = df[df["route"] == "phase2"]

# Average trust score by modality
df.groupby("modality")["final_score"].mean().sort_values()