File size: 2,762 Bytes
fb56533 0ffea57 fb56533 16eaf62 fb56533 0ffea57 fb56533 0ffea57 fb56533 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 | ---
license: odc-by
task_categories:
- text-generation
- question-answering
language:
- en
tags:
- synthetic
- sft
- instruction-tuning
- education
size_categories:
- 100K<n<1M
configs:
- config_name: math
data_files: math.parquet
- config_name: general_knowledge
data_files: general_knowledge.parquet
---
# LittleCurriculum-Chat
Synthetic K–5 chat data generated with Gemini 2.5 Flash and filtered with the
[LittleCurriculum filter](https://github.com/littlelearner-ll/littlecurriculum-filter).
Used to train the [LittleLearner](https://huggingface.co/littlelearner) models.
| Config | Rows | Seeded from | Content |
|---|---|---|---|
| `math` | 79,543 | MegaMath-Web-Pro-Max | Word problems with step-by-step worked solutions |
| `general_knowledge` | 484,787 | [LittleCurriculum](https://huggingface.co/datasets/littlelearner/LittleCurriculum) | Reading comprehension, factual QA, explanation, summarisation, definitions |
Seeds are real documents rather than topic prompts, which keeps the topic
distribution broad. The seed documents are not included here, so some questions
assume context the reader does not have ("what is the main idea of this
article?"). `seed_id` is the FineWeb-Edu document id, so the source text can be
recovered by joining against
[LittleCurriculum](https://huggingface.co/datasets/littlelearner/LittleCurriculum).
Every row carries a `messages` column in `[{role, content}]` form alongside the
structured fields.
```python
from datasets import load_dataset
ds = load_dataset("littlelearner/LittleCurriculum-Chat", "general_knowledge")
```
## Related data
For instruction fine-tuning we additionally used SmolTalk, MMLU
auxiliary-train and GSM8K, each filtered with our pipeline. For GRPO we used
grade-stratified synthetic problems from the MathCAMPS pipeline alongside
GSM8K, again filtered with our pipeline. We do not redistribute these — they
derive from public datasets and reproduce in one command:
```bash
python filter_k5.py --hf-dataset HuggingFaceTB/smol-smoltalk --out smoltalk_k5.parquet
```
In our runs this data was most useful mixed into pretraining and midtraining
rather than reserved for a dedicated SFT stage.
## License
ODC-By 1.0, inherited from FineWeb-Edu via LittleCurriculum. Generated with
Google Gemini.
## Citation
```bibtex
@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},
}
```
|