Datasets:

littlelearner
/

LittleCurriculum

@@ -13,129 +13,47 @@ tags:
 - education
 - fineweb-edu
 - knowledge-boundary
-- common-core
 ---
 # LittleCurriculum
-> ## 🚧 Upload in progress
-> This corpus is **still uploading** — not all shards are present yet, so the dataset is
-> **incomplete**. Please **do not download or benchmark** until this notice is removed.
-> Expected complete: within a few hours. Thanks for your patience.
-**LittleCurriculum** is a **developmentally restricted** ≈88-billion-token English pretraining corpus:
-a **precision-first** subset of [FineWeb-Edu](https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu)
-filtered to the U.S. **K–5** developmental level. A multi-stage pipeline enforces constraints on
-**knowledge, syntactic complexity, and reasoning** (with Common Core State Standards as the grade-level
-reference), removing material characteristic of the **Beyond-K–5** curriculum—advanced concepts such as
-algebra or higher-grade science—while retaining text a K–5 learner can process.
-It is the corpus behind the **LittleLearner** models — language models trained from scratch under a
-*pedagogically controlled knowledge boundary*, released alongside matched unfiltered controls. See the
-[LittleLearner model collection](https://huggingface.co/littlelearner).
-> **A note on scope.** The pipeline constrains the **developmental level** of a text—the knowledge,
-> linguistic complexity, and reasoning it demands—not its subject matter. Advanced **Beyond-K–5 concepts**
-> (calculus, quantum physics, and the like) are filtered to near-zero, but subject matter itself is not
-> curated: everyday content on any topic a curious child might follow—aviation, current events, crime in
-> the news—can appear when written at a K–5 level. The design is deliberately *precision-first*, trading
-> recall for a sharp, interpretable boundary (validated on CommonCoreText and WeeBit; see paper §3.1).
-> See **Scope of filtering** and **Limitations** below.
-## Dataset at a glance
-| | |
-|---|---|
-| Tokens | ≈88B (measured ≈86–88B; tokenizer-dependent) |
-| Documents | ~242M |
-| Shards | 2,546 Parquet files (`shard_00000.parquet` … `shard_02545.parquet`) |
-| Compression | ZSTD |
-| Schema | `id` (string), `text` (string) |
-| Language | English |
-| Source | FineWeb-Edu (CommonCrawl-derived) |
-| License | ODC-By 1.0 |
-*Token counts were measured with the LittleLearner tokenizer (vocab 32,768; ≈4.43 chars/token). The
-paper reports the corpus as "88B tokens"; the LittleLearner models were each trained on **80B tokens**
-(≈0.9 epoch of this corpus) — the corpus size and the training budget are distinct quantities.*
-## Provenance
-```
-CommonCrawl  →  FineWeb-Edu  →  [5-stage K–5 filter]  →  LittleCurriculum
-```
-## Filtering pipeline
-The corpus is produced by applying five sequential filters to FineWeb-Edu documents. The reusable,
-self-contained implementation is released as supplementary material with the paper.
-1. **Age-of-Acquisition + Zipf.** Drop a document if its out-of-vocabulary fraction exceeds 5%, or if
-   the 95th-percentile word Age-of-Acquisition exceeds 12 years. Unknown words are imputed an AoA from
-   Zipf word frequency.
-2. **FastText grade classifier.** Keep a document only if a 4-class (K5 / K8 / K12 / out-of-scope)
-   fastText classifier assigns it the **K5** label.
-3. **ModernBERT grade classifier.** Keep a document only if a fine-tuned ModernBERT sequence classifier
-   assigns it the **K5** class (independent second classifier; documents must pass *both*).
-4. **Symbolic filter.** Drop documents matching any of ~28 advanced-mathematics / notation patterns
-   (exponents, radicals, integrals/sums, `\frac`, function notation, chemical formulas, etc.).
-5. **Word-list filter.** Drop documents containing any word from a corpus-derived blocklist of terms
-   strongly associated with beyond-K–5 material (log-odds `delta_k5 ≤ -4`).
-## Intended use
-Research on **controlled knowledge exposure**: studying whether a capability was *learned* from the
-training distribution or merely *elicited* by post-training/scaling/prompting, when the training
-distribution's coverage is explicitly specified. Suitable for pretraining language models with an
-interpretable knowledge boundary, continual-learning and boundary-probing studies, and machine-vs-child
-learning comparisons.
-## Scope of filtering: developmental level, not topic
-Consistent with the paper, the pipeline enforces constraints on **knowledge, syntactic complexity, and
-reasoning**—the *developmental level* a text demands—rather than curating subject matter. Two things
-follow: (i) advanced **Beyond-K–5 concepts** (e.g. calculus, quantum physics) are removed to near-zero;
-and (ii) subject matter is otherwise not restricted, so everyday content on any topic a curious child
-might follow—aviation, current events, crime in the news—can appear when written at a K–5 level. The
-filter is *precision-first*: tuned for near-zero Beyond-K–5 retention at the cost of discarding much
-in-scope material, yielding a sharp, interpretable boundary rather than an exhaustive sample of K–5 text.
-## Limitations
-- **Developmental-level filter, not topical curation.** See **Scope of filtering** above — the pipeline
-  removes Beyond-K–5 *concepts*, not general subject matter, which can span any domain at a K–5 level.
-- **Not zero-exposure.** The filter removes advanced academic *content and skills*, not every token. A
-  measured spot-check found beyond-curriculum vocabulary present only at low, grade-appropriate
-  frequency (e.g. "adrenaline" 0.075%, "hormone" 0.177%, "epinephrine" 0.006%, "ballistic missile"
-  0.001% of documents), overwhelmingly in everyday registers (news/history, food labels, first-aid,
-  narrative) rather than advanced technical instruction. The capability-boundary claim concerns
-  *skills*, not *lexical exposure*.
-- **Evaluation scope.** In the accompanying paper, the "elicitation, not acquisition" finding is
-  evaluated primarily on **mathematical reasoning** (MathCAMPS, accuracy by grade). General-knowledge
-  boundary behavior is characterized qualitatively via targeted probes rather than a systematic
-  benchmark.
-- **Inherited biases.** As a filtered subset of FineWeb-Edu (web text), the corpus inherits the
-  coverage, quality, and biases of that source and of CommonCrawl.
-- **Classifier imperfection.** The grade classifiers and heuristics are imperfect; both false drops
-  (in-scope text removed) and false keeps (borderline text retained) occur.
-## License & attribution
-Released under the **Open Data Commons Attribution License (ODC-By) v1.0**, inherited from FineWeb-Edu.
-Use requires attribution to this dataset and to the upstream source. Underlying text originates from
-CommonCrawl; please also observe the CommonCrawl terms of use.
 ## Citation
 ```bibtex
-@article{littlelearner2026,
-  title   = {LittleLearner: Language Models Under Pedagogically-Controlled Knowledge Exposure},
-  author  = {Fanfei Li and Jana Zeller and Manuel Prada-Corral and Thadd{\"a}us Wiedemer
-             and Prasanna Mayilvahanan and Ryan Cotterell and Wieland Brendel},
-  journal = {arXiv preprint arXiv:2608.13545},
-  year    = {2026},
-  url     = {https://arxiv.org/abs/2608.13545}
 }
 ```
-Please also cite **FineWeb-Edu** (Penedo et al., 2024) as the upstream source.

 - education
 - fineweb-edu
 - knowledge-boundary
+- common-core-standards
 ---
 # LittleCurriculum
+**🚧 Upload in progress, do not benchmark yet!**
+**LittleCurriculum** is an ~88B-token English pretraining corpus derived from [FineWeb-Edu](https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu-score-2). It is filtered to align with U.S. Common Core standards for grades **K–5**, removing documents containing academic concepts and skills characteristic of later grades.
+It is the training corpus for the **LittleLearner** models, designed to study language models under a controlled knowledge boundary. See the [LittleLearner paper](https://arxiv.org/abs/2608.13545) for methodology, validation, and experiments.
+## Dataset
+* **Documents:** ~242M
+* **Tokens:** ~88B
+* **Language:** English
+* **Source:** FineWeb-Edu
+* **Format:** Parquet (`id`, `text`)
+* **License:** ODC-By 1.0
+## Dataset Construction
+LittleCurriculum is produced from FineWeb-Edu using five sequential filters:
+1. Age-of-Acquisition and word-frequency filtering
+2. fastText grade-level classification
+3. ModernBERT grade-level classification
+4. Advanced mathematical/symbolic notation filtering
+5. Beyond-K–5 vocabulary filtering
+The dataset constrains **academic curriculum content**, not all lexical or real-world exposure. Words associated with advanced topics may still occur in everyday, grade-appropriate contexts. Full construction details, thresholds, validation, and limitations are described in the [paper](https://arxiv.org/abs/2608.13545).
 ## Citation
 ```bibtex
+@misc{li2026littlelearnerlanguagemodelspedagogically,
+      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},
 }
 ```