littlecurriculum-filter

LittleCurriculum Filter — Artifacts

Classifier weights and lexical resources for the LittleCurriculum five-stage K–5 text filter.

These files are the runtime dependencies of littlelearner-ll/littlecurriculum-filter.

Usage

git clone https://github.com/littlelearner-ll/littlecurriculum-filter
cd littlecurriculum-filter
pip install -r requirements.txt
python download_artifacts.py          # fetches this repo
python filter_k5.py --in shard.parquet --out kept.parquet

Contents

Both classifiers predict one of K5, K8, K12, OOS (id2label = {0: "K5", 1: "K8", 2: "K12", 3: "OOS"}); the filter retains documents predicted K5.

Training

The classifiers are distilled from LLM-as-a-judge annotations of FineWeb-Edu, generated with Google Gemini using prompts initialised from the Common Core State Standards and refined with automatic prompt optimisation. Full annotation of FineWeb-Edu would have been prohibitively expensive, which is what motivates the cascaded design: a cheap lexical stage, then fastText, then the ~50× more expensive ModernBERT.

Intended Usage

Trained for web prose. Labels come from FineWeb-Edu documents. On substantially different distributions, retraining is recommended.

Fixed to the K–5 boundary. Retargeting to a different grade band requires retraining these classifiers.

Expect whole documents. The classifiers estimate a document's overall grade level and have little to work with in a single sentence. Application to full documents is recommended.

Citation

@misc{li2026littlelearner,
      title={LittleLearner: Language Models Under Pedagogically Controlled Knowledge Exposure},
      author={Fanfei Li and Jana Zeller and Manuel Prada-Corral and Thaddäus Wiedemer and Prasanna Mayilvahanan and Ryan Cotterell and Wieland Brendel},
      year={2026},
      eprint={2608.13545},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2608.13545},
}

Age-of-Acquisition norms: Kuperman, Stadthagen-Gonzalez & Brysbaert (2012), Age-of-acquisition ratings for 30,000 English words, Behavior Research Methods 44(4).

License

Apache-2.0, matching the ModernBERT base model and fastText.