Yurai

yurai — 由来, "origin; where a thing comes from."

Yurai is a provenance auditor for AI models. It traces model lineage, audits license inheritance, and flags trust gaps across Hugging Face models.

Run it from the CLI, wire it into CI, or explore findings in an investigation UI.

Know what you're depending on.

Live demo · Install

Findings

• License inheritance violations: Apache-2.0 declared on a Llama derivative that's actually governed by Meta's Community License
• Transitive license violations: permissive license on a model whose grandparent or earlier ancestor uses a copyleft or restricted license
• Lineage inconsistencies: declared base model doesn't match the architecture in config.json
• Gated-derivative detection: public models derived from gated parents, bypassing access controls (uses the direct HF gated field, falls back to license heuristics)
• Documentation gaps: missing license or base model declarations
• Trust signals: new uploader accounts, zero community engagement, high downloads with zero likes, recently modified old models
• Metadata anomalies: weight sizes that don't match the declared architecture, undeclared quantization, suspicious or missing weight files

Each finding includes a severity, a reason explaining why it matters, and the raw declared-vs-actual values that triggered it.

Install

cargo install yurai

Usage

# Investigate a model
yurai investigate meta-llama/Llama-3.1-8B-Instruct

# JSON output
yurai investigate ruslanmv/Medical-Llama3-8B --json

# SARIF output (for GitHub code scanning)
yurai investigate ruslanmv/Medical-Llama3-8B --sarif

# Fail CI on high-severity findings
yurai investigate some/model --fail-on-high

Batch mode - investigate multiple models from a file or stdin:

# From a file
yurai batch --from models.txt

# From stdin
echo -e "microsoft/phi-2\nruslanmv/Medical-Llama3-8B" | yurai batch

# Batch with SARIF output
yurai batch --from models.txt --sarif results.sarif

Set HF_TOKEN to access gated models:

export HF_TOKEN=hf_...
yurai investigate meta-llama/Llama-3.1-8B-Instruct

Web Explorer

Three-panel investigation UI: lineage graph, tabbed evidence details, and findings with declared-vs-actual diffs. Click a finding to highlight the related evidence across all panels.

gavxm.github.io/yurai

Run locally:

cd web && npm install && npm run dev

GitHub Action

Add provenance checks to your CI pipeline:

- uses: gavxm/yurai@v0.3.0
  with:
    models: |
      meta-llama/Llama-3.1-8B-Instruct
      ruslanmv/Medical-Llama3-8B
    fail-on-high: true
    hf-token: ${{ secrets.HF_TOKEN }}

The Action investigates each model and posts a summary to the job output. Set fail-on-high: true to block merges when HIGH severity findings exist.

How It Works

Yurai fetches evidence from four HuggingFace sources concurrently, then runs cross-referenced checks across them:

Source	What it provides
HF metadata	license, base model, tags, downloads, likes, gated status, file listing, timestamps
Model tree	multi-hop lineage chain (up to 4 ancestors), licenses, gated status, siblings
config.json + safetensors	architecture, parameters, weight size, quantization config
Community signals	uploader account age, discussion activity

The key insight is gap-as-signal: contradictions between sources are the findings, not incidental noise.

Architecture

src/lib.rs       :types, public API, schema
src/engine.rs    :investigation orchestration
src/main.rs      :CLI, batch mode, SARIF output
src/render.rs    :terminal text rendering
src/sources/     :evidence fetchers (HF metadata, model tree, config, community)
src/findings/    :cross-referenced checks (license, lineage, gated, trust, metadata, doc gaps)
web/             :React + Vite + Tailwind

License

AGPL-3.0. See LICENSE.