ModelLens — Early Tester Guide

ModelLens lets you upload a trained machine learning model and instantly explore how it behaves — which features drive its decisions, where it might fail, and what it predicts for specific inputs — all through a conversational interface.

The reasoning engine behind ModelLens is Perun. Perun uses the weight2text approach: it studies how neural networks learn from large collections of expert models and translates that understanding into natural-language explanations.

You do not need to modify your model, retrain anything, or write any code. Upload the file, wait for exploration to complete, then start asking questions.

Supported formats

• .pt / .pth — PyTorch nn.Module (saved with torch.save)
• .keras / .h5 — Keras / TensorFlow model
• .pkl / .joblib — scikit-learn estimator

Required inputs

• Feature names — comma-separated list matching your model's input order, e.g. age, income, credit_score
• Task type — classification or regression
• Class names — classification only, e.g. fraud, legitimate

Optional inputs

• Scaler file — if you applied a StandardScaler or MinMaxScaler before your model, upload the fitted scaler as a .pkl or .joblib file. ModelLens will apply it automatically before every prediction.
• Feature bounds — min/max per feature. If omitted, ModelLens infers them from the scaler or uses defaults.
• Sample CSV — a small CSV of real data rows (no label column). Used to show example predictions and improve exploration quality.

After upload, ModelLens automatically runs an exploration step. This probes your model across its feature space to build a Knowledge Card — a structured summary that Perun reads before answering any question.

Exploration typically takes 30–90 seconds depending on model complexity. The status is shown on the model card: exploring → complete. Chat is locked until exploration finishes.

What exploration computes:

• Decision boundary samples across the feature space
• Per-class optimal and boundary-case inputs
• Rough feature sensitivity at multiple operating points
• Architecture summary (for neural networks)

Once exploration is complete, the chat panel unlocks. Perun has read the Knowledge Card and can answer questions about your specific model.

Good questions to start with

• "Which features matter most?"
• "What does the model predict for a typical high-risk case?"
• "Show me how the prediction changes as [feature] increases"
• "What are the decision boundaries between [class A] and [class B]?"
• "Find edge cases — inputs near the decision boundary"
• "Plot [feature A] vs [feature B] coloured by predicted class"
• "Why did it predict [class] for this sample?" — use the sliders to set a specific input first

Charts

Perun can generate scatter plots, bar charts, and line charts inline in the chat. Just ask — for example: "Plot a bar chart of feature importance" or "Show a scatter of age vs income coloured by prediction."

Session continuity

Your conversation history is saved per model. Refreshing the page or returning later will restore the previous conversation.

For classification models, click the bar chart icon in the top-right of the model page to open Deep Analysis.

Deep Analysis runs a more thorough offline computation (1–5 minutes) and produces:

• Permutation feature importance — how much each feature affects predictions when shuffled
• Scenario analysis — representative operating points across the feature space, with predicted class and confidence
• Per-class feature profiles — typical feature values for each predicted class

On the left side of the chat panel (or below on mobile), you will find a set of feature sliders. These let you set specific input values and see the model's live prediction.

Use the sliders to:

• Construct a specific sample and ask Perun to explain the prediction
• Explore how the prediction changes as you move one feature at a time
• Find the exact threshold where the model flips from one class to another

The prediction updates in real time as you adjust the sliders.

Your feedback is the most valuable thing during the early access phase. There are several ways to give it:

• Like / Dislike buttons — appear below the latest Perun response. Thumbs up if the answer was useful; thumbs down if it was wrong, vague, or unhelpful.
• Retry — re-runs the same question with a fresh response. Useful if Perun got confused mid-conversation.
• Email — for longer feedback, bugs, or feature requests, email info@zerooneresearch.ai.

• PyTorch models with custom layers — models that use non-standard layers or require special forward arguments may fail to load. Try saving with torch.jit.script or contact us.
• Very high-dimensional inputs — models with more than ~100 features will explore and analyse correctly but Perun's explanations may be less precise.
• Regression models — deep analysis is not yet available. Chat and the prediction panel work fully.
• Extrapolation warnings — if you move sliders outside the model's training range, Perun will flag this in its response. The prediction is shown but may be unreliable.
• Chat message limit — free accounts have 50 chat messages per month. The counter resets on the 1st of each month and is shown in your profile popover.
• Model limit — free accounts can upload up to 3 models. The 100+ demo models in the catalogue do not count against this limit.

We read every message during the early access phase.

• General feedback & bugs: info@zerooneresearch.ai

← Back to home