BankBench + Mechanistic Interpretability

Gemma 2 9B IT — Safety Eval Dashboard

Fetching experiment data...

BankBench 20-Scenario Eval

Gemma 2 9B IT — Safety Results

Safety Scores by Scenario

Pass/Fail by Category

Action Distribution

Multi-Turn Depth vs Safety

Per-Scenario Detail

All 20 Scenarios — Turns, Responses & Mech Interp

Click any scenario to expand full turn-by-turn responses and per-scenario mechanistic interpretability.

Mechanistic Interpretability

Internal Representations — All Scenarios

Aggregate analysis of hidden state norms, safety direction probes, and activation patterns.

Hidden State Norms by Layer

Norm Deltas (Decision Points)

Safety Direction Probe

Logit Lens — First Scenario

Eval Awareness Detection

Does Gemma 2 Know It's Being Tested?

Comparing internal representations when evaluation cues are present vs absent.

Anomaly Detection

Flagged Anomalies & Findings

Automatically detected patterns that warrant investigation.

Capabilities Eval

Banking Tool Use — Gemma 2 vs Mistral 7B

Correct tool selection, argument accuracy, and clarification quality across 20 customer interaction scenarios.

Capability Scores by Scenario

Model Comparison

Score Breakdown by Dimension

Straightforward vs Edge Cases

Per-Scenario Results