When Nemotron thinks it's Qwen
I asked a 4B NVIDIA model "tell me about yourself", and it confidently introduced itself as a different company's model.
#local-llms#identity#evals
NVIDIA’s Nemotron 3 Nano 4B is a compact hybrid Mamba2-Transformer reasoning model, compressed from Nemotron Nano 9B v2 and built for edge inference. It’s a good model.
I asked it the simplest possible question: “Tell me about yourself.”
It said it was Qwen, a model made by a different lab.
Why it happens
Compressed and distilled models learn from large amounts of synthetic data, and some of that data is other models talking about themselves. Unless the identity is reinforced, for example in a system prompt or in post-training, the model repeats whichever self-description it saw most often.
Why it matters
- Identity is a behavior, not a fact the model knows. If your product depends on what the model says it is, test that the same way you test any other output.
- Always ship a system prompt, even with local models. The default persona is whatever the training data happened to contain.
- This is exactly the kind of check Tribunal formalizes: identity is one of its six graded dimensions.