Is your model smarter than a 5th grader?

# Is your model smarter than a 5th grader? I saw a post about [LittleLearner](https://littlelearner-ll.github.io/), a scale-up study where Qwen3 dense models are pretrained on a corpus filtered down to U.S. K-5 material ([paper](https://arxiv.org/abs/2608.13545)). This model should thus effectively be a 5th grader. If you were watching TV in the early 2000s, you've most likely seen this show: It puts an adult in front of grade-school questions with a real fifth grader sitting next to them. So I ran the [`littlelearner/littlelearner-5b-chatty`](https://huggingface.co/littlelearner/littlelearner-5b-chatty) checkpoint against it, just to see what would come out. ## Getting the questions There is no usable dataset for this show. The subtitle sites do not have it all, and the episodes that are officially online are a small slice of the run. I limited myself to the first three seasons, where three sources together cover everything: | Source | Episodes covered | | --- | --- | | [GSNN's…