Working…

Traceix Traceix
IPFS Datasets

AI Security Lab

A guided lab to peer review AI training datasets. You’ll pick a dataset release (TUUID), fetch 2–5 samples, compare “what the AI saw”, and export a peer review report you can share or archive.

Keep labs free → donate
What am I looking at?
Close

Each sample record includes normal metadata + a field named decrypted_training_data. That training data is a structured set of features (numbers/flags/fields) used by the model.

This lab helps you verify: (1) schema consistency, (2) sane numeric ranges, (3) label plausibility, (4) duplicates/outliers, and (5) provenance fields.

Progress
Tip: Use the big buttons. If something is locked, it will tell you exactly what to do next.
Peer Review Steps
Do these in order. The big Next button stays disabled until you're ready.
Current step: 1 • Pick a dataset to begin.
Step 1 — Pick a dataset release (TUUID)
TUUIDs load automatically. Choose one and we’ll load its hash list.
Quick start
1) Pick a dataset in the dropdown • 2) Click Load dataset • 3) Click Next
What is a TUUID? Close

A TUUID is a dataset release ID. Each release contains a list of sample hashes. You pick a release, then you pick a few sample hashes from it to review.

Your goal is to sanity-check what’s inside: consistent schema, plausible values, and label/provenance quality.

Step 2 — Pick 2–5 sample hashes
Use Random 2 if you don’t know what to pick.
Hashes in this dataset
—
Selected: 0 / 5
Minimum is 2.
How to choose samples Close

Easy mode: click Random 2 and move on.

Better: pick one “normal looking” and one “odd looking” sample after you fetch and see values.

Finding: if multiple hashes map to the same CID later, that’s worth noting in the report.

Step 3 — Fetch sample records
For each hash: resolve CID → fetch dataset record. Retries on 429/5xx.
Pipeline per hash: ipfs/find → ipfs/search.
What should I watch for? Close

After fetch, you’ll review:

  • Missing fields / inconsistent schema
  • Outliers (entropy, counts, sizes)
  • Verdict vs capabilities mismatch
  • Duplicates (same CID across hashes)
Quests
Tiny tasks that teach peer review.
Score: 0
Fetched samples
Add notes per SHA. Then choose left/right and click Compare.
SHA-256 CID Verdict Model Capabilities Notes
Comparison (with explanations)
Tabs show different “views” of the same training features. Use the explanations to learn what each view means.
What “Vector + keys” means
The model can’t read “files” directly. It uses a list of numeric features (a vector). Here you can see which keys were used and what values went into the vector.
Left
—
Verdict
—
File type
—
Feature preview 
Select and compare two samples.
Capabilities
—
Right
—
Verdict
—
File type
—
Feature preview 
Select and compare two samples.
Capabilities
—
Key differences
Shows numeric fields with the largest differences. Use this to spot outliers fast.
Showing —
Feature Left Right Δ Why it matters
What “Core metrics” means
These are common structural fields (sizes, counts). Big anomalies often indicate packing, corruption, or unusual builds.
Left snapshot
—
Compare two samples to render.
Right snapshot
—
Compare two samples to render.
Side-by-side bars
Bars scale per metric (max(left,right)).
Compare two samples to render.
What “Entropy” means
Entropy (0–8-ish) tends to rise with packing/compression/encryption. Spikes in sections/resources are common review targets.
Entropy-focused bars
Compare two samples to render.
What “Imports/exports” means
Very low imports can indicate packing/stub loaders. Very high counts can indicate heavy linkage or unusual compilation.
Imports/exports + structure
Compare two samples to render.
What “Capabilities” means
Capabilities are extracted behaviors/traits. If the model verdict and capabilities strongly disagree, that’s a strong peer-review note.
Overlap summary
Intersection vs unique capabilities.
Compare two samples to render.
Unique capabilities
What appears on one side only.
Compare two samples to render.
Notes reminder
Add notes in the samples table above (per SHA). Notes export into your report.
Step 5 unlocks after you run a comparison. You still move forward only when you click Next.
Step 5 — Export + build your own tooling
Export your report and copy starter code (Python + JS) for your own parser.
Python: fetch + vectorize

                    


                    

                        

                            
JavaScript: fetch + pacing

                            
                        

                        

                    


                    

                        
Peer review checklist

                        
    
                            
  • • Feature keys consistent across samples? Missing keys can bias a model.
    • 
                            
  • • Numeric ranges sane? (entropy typically 0–8-ish, counts not absurdly high)
    • 
                            
  • • Verdict plausible vs capabilities?
    • 
                            
  • • Duplicates/near-duplicates (same CID / repeated hash)?
    • 
                            
  • • Provenance present (license, upload timestamp, model version, tx where relevant)?
    • 
                        


                        

                            
                            Exports a JSON summary of your selection, fetched records, notes, and compare settings.