Inspecting a Dataset

The Dataset Inspector analyses an HDF5 dataset stored in RIA Hub and surfaces statistics and quality indicators across several dimensions. Analyses run as background jobs and results are returned as tables and summary metrics directly in the browser.

Use the Inspector when you want to:

• Verify class balance before training — unequal sample counts per class can bias a model toward over-represented modulations
• Compute per-class signal metrics (RMS, power, PAPR, SNR, EVM) to spot normalisation issues or outlier classes
• Flag anomalous samples using statistical methods (Z-score, IQR, or MAD) before they corrupt training
• Compare two datasets side by side to check consistency between versions or independently generated splits
• Diff two dataset versions to understand what changed between curation runs

Analysis modes available:

Mode	What it computes
Overview	File metadata, LFS ID, and creation attributes — no computation required
Class Balance	Sample counts per class, balance ratio, imbalance severity
Per-Class Statistics	RMS, power, PAPR, SNR, EVM, crest factor, bandwidth per modulation class
Anomaly Detection	Outlier samples flagged by Z-score, IQR, or MAD with configurable threshold
Comparison	Side-by-side metrics across two or more datasets
Diff	What changed between two versions of the same dataset

What you’ll need

• An HDF5 dataset stored in a RIA Hub repository with Git LFS enabled — the inspector looks up files by their LFS object ID
• Larger datasets take longer per analysis; expect a few seconds to a minute depending on dataset size

Step 1 — Open the Inspector

Navigate to your repository, click the Dataset Manager tab, then select Inspector from the sidebar.

Step 2 — Select a repository and dataset

Use the repository dropdown to find your repo (you can type to search). Then pick the HDF5 file from the dataset list below it.

Step 3 — Review the overview

The Overview section shows metadata pulled directly from the file without running any computation:

• Repository location, branch, and commit hash
• LFS object ID
• File-level attributes written when the dataset was created (e.g., sample rate, number of classes)

This is a quick sanity check — confirm the right file and version are selected before running any analyses.

Step 4 — Check class balance

Open the Class Balance section and click Run Analysis.

The inspector counts the samples belonging to each modulation class and reports:

• Total sample count
• Number of classes
• Per-class sample counts
• Balance ratio and imbalance severity

A perfectly balanced dataset has equal samples per class. When some classes have far more samples than others, models trained on this data may learn to favor the over-represented classes.

Tip

Aim for a balance ratio close to 1.0. If your dataset is severely imbalanced, consider re-curating with the Balanced slicer strategy, or generating a new synthetic dataset using the Balanced sampling strategy in the Generator.

Step 5 — View per-class statistics

The Per-Class section computes signal metrics for each modulation class. Click Run Analysis to start.

Metric	What it measures
RMS	Root-mean-square amplitude — a measure of average signal power
Power	Mean squared amplitude
PAPR	Peak-to-Average Power Ratio — how “peaky” the waveform is
SNR	Estimated signal-to-noise ratio
EVM	Error Vector Magnitude — how closely symbols match ideal constellation points
Crest factor	Ratio of peak amplitude to RMS
Bandwidth	Estimated occupied bandwidth

Results appear as a table with one row per class. Look for outliers — a class with a much higher RMS than the others, for example, may indicate a normalization issue in your data.

Note

SNR and EVM estimates are most meaningful when your dataset was generated synthetically with known noise levels. For real-world recordings they are approximations.

Step 6 — Detect anomalies (optional)

The Anomaly Detection section flags individual samples that look unusual compared to the rest of the dataset.

Choose a detection method:

Method	How it works
Z-score	Flags samples more than N standard deviations from the mean
IQR	Uses the interquartile range to define “normal” — more robust for non-Gaussian distributions
MAD	Median Absolute Deviation — the most robust option when the data contains extreme outliers

Set a threshold (higher values flag fewer samples) and click Run Analysis. Flagged samples may be corrupted, mislabeled, or otherwise unusual — worth investigating before training.

Step 7 — Compare datasets (optional)

The Comparison section lets you compare two or more datasets side by side. This is useful when:

• Comparing a freshly curated dataset to its source recordings
• Verifying that two independently generated datasets share similar statistical properties
• Checking that a new dataset version is consistent with an older one

Select the datasets to compare, choose which metrics to include, and run the comparison.

Step 8 — Diff two dataset versions (optional)

The Diff section shows what changed between two versions of a dataset. Use this when you’ve re-curated or re-generated a dataset and want a systematic account of what’s different.

---

Next steps

• Train a model — When class balance and signal statistics look good, take the dataset to the Model Builder to launch a training run.
• Re-curate if needed — If the dataset has class imbalance or quality issues, return to the Curator to adjust the slicer, qualifier, or recording selection.