TL;DR: AI bias detection here is a paired A/B audit. Each scenario varies exactly one demographic parameter, every answer is scored against a fixed rubric, and the per-axis breakdown shows where bias concentrates instead of one vague score. On the Bias Stereotypes leaderboard, HY3 (free) leads at 82.9 and Nemotron 3 Ultra (free) follows at 81.7.
Most bias scores are a single number. That number tells you a model is biased, not where. This benchmark is built the other way around.
1. Paired A/B scenarios. Each scenario varies exactly one demographic parameter: name, gender, class, origin, city, or political sensitivity. Everything else stays fixed. That is the control.
2. Probing on stereotype axes. Models are probed across axes like cultural bias, language bias, double standard, evaluation bias, intersectionality, default generation, and factual neutrality. Each axis is a percentage from 0 to 100, and higher is better.
3. Deterministic symmetry scoring. The core fairness signal is response symmetry. If changing only the name flips the tone, the pair is asymmetric. That is measured deterministically, not by vibes.
4. A fixed LLM judge. A judge model scores the same outputs against a fixed rubric. The judge model is recorded as its own field, so you can see what did the judging.
5. Per-axis breakdown. The output is not one score. It is a profile. A model can be strong on factual neutrality and weak on intersectionality, and the breakdown shows exactly that.
| Model | Score | Provider |
|---|---|---|
| HY3 (free) | 82.9 | opencode-zen |
| Nemotron 3 Ultra (free) | 81.7 | opencode-zen |
We also submitted our own model to this benchmark. GLM 5.2 scored 78.0 overall.
A 1.2 point gap between first and second is small. In a symmetry-based audit, that usually means both models handle the obvious cases well: they do not flip tone when a name changes, and they do not collapse on a single demographic swap.
The 78.0 for GLM 5.2 is the more interesting number. It is not a failure. It is a signal that at least one axis is dragging the average down. In practice, a model at 78.0 is often strong on factual neutrality and weaker on intersectionality or double standard, where two demographic parameters interact and the model has less training signal to fall back on.
That is the point of the per-axis view. A single 78.0 would tell you nothing actionable. A per-axis profile tells you which prompt families to harden.
If you are choosing a model for a user-facing product, do not pick on the overall score alone. Pull the axis breakdown and match it to your risk surface. The leaderboard is a starting filter, not a verdict. The 82.9 and 81.7 models are close enough that your own prompt distribution should decide the tie.
This audit measures response symmetry on paired scenarios. It does not measure downstream harm, real user outcomes, or bias in retrieval and tool use. The judge is an LLM, and LLM judges have their own biases, which is why the judge model is recorded as a field. A high score means the model is symmetric on these axes. It does not mean the model is fair in every context.
AI bias detection works best when it is decomposed. Probe on stereotype axes, score against a fixed rubric, and report the per-axis breakdown so you can see where bias concentrates. Start with the full leaderboard at lforla.org, then run the audit on your own prompts.