Team: Richard Tibbles, Jamie Alexandre
Overview: Science misconception research spans decades but isn’t operationalized for classroom use. Three public goods, all openly licensed: a benchmark for identifying science misconceptions in free-form student responses, a structured dataset, and reference classifiers for classroom devices.
Subject Focus: Grades 6-12 Earth, Life, and Physical sciences
Targeted Universalism Focus: The targeted strategy focuses directly on the economic and infrastructure barriers characteristic of under-resourced schools by building and evaluating AI models designed to run entirely on the constrained hardware profiles these schools actually possess (such as Chromebooks, Android phones, Raspberry Pi units, and older laptop CPUs).
Public Goods & Deliverables:
- A collection of 100 to 200 open assessment packages containing free-response prompts, comprehensive scoring rubrics, citations, and standards alignments mapped via the CZI Knowledge Graph (CC-BY-4.0).
- Evaluation Benchmark: A standardized classification evaluation suite released publicly on Hugging Face to test external classifiers.
- Fine-tuned, parameter-efficient small language model reference baselines (such as sub-3B parameter Llama, Qwen, or Phi-mini variants) quantized into 4-bit and 8-bit versions for private, offline execution.