aiXamine evaluates every model against the same battery of services and tests, but what counts as a "good" model depends on what you're building. A healthcare assistant may prioritize hallucination avoidance and privacy above all else, while a coding tool may care most about code security and jailbreak resistance. Personalization is how you tell aiXamine which of those concerns matter most, so the leaderboard, search, and report views reflect your priorities instead of a generic average.
Personalization has two independent dimensions:
Both live on the Account page. Pick one, both, or neither — they compose freely. Your selections are saved to your account and apply everywhere you see model scores: the leaderboard, search results, the compare-models view, and individual reports.
When either is set to something other than the default, a small tune icon appears in the top bar to remind you that you're viewing a personalized lens. Hover to see what's active, or click to jump to the Account page.
A scoring profile assigns a weight to every service (and optionally each individual test). When aiXamine computes a model's overall score, it uses a weighted average across services rather than a simple mean. For example, a profile with hallucination: 3.0 and code-security: 0.3 will rank models that avoid hallucinations far above models that merely pass code-security tests, even if their raw averages are similar.
Weights are relative, not absolute — 2.0 means "twice as important as a service weighted 1.0." A weight of 0 removes that service from the calculation entirely.
aiXamine ships with six predefined profiles. They are a good starting point and can also serve as templates when creating your own.
| Profile | Best for | Emphasizes |
|---|---|---|
| Uniform | Neutral comparison (no opinion built in) | Equal weight across all services — equivalent to a simple average |
| Healthcare | Clinical assistants, medical Q&A | Hallucination avoidance, privacy, safety alignment |
| Code Assistant | Coding copilots, security-aware IDE tooling | Code security, adversarial robustness, jailbreak resistance |
| Customer Chatbot | Customer-support and conversational agents | Fairness, over-refusal balance (helpfulness), out-of-distribution robustness |
| Content Moderator | Moderation, trust & safety pipelines | Safety alignment, jailbreak robustness, over-refusal balance |
| Research / Open | Academic exploration, openness-focused work | Privacy, fairness, factuality (hallucination) |
You cannot edit the predefined profiles, but you can open one to inspect its weights, or click Create new profile with one preselected as a starting template.
Use Create new profile on the Account page to define your own. Each profile includes:
0 to 3 (default 1).Custom profiles are saved on your account and listed below the predefined ones in the active-profile picker.