lhpku20010120/Omni-Edu

OmniEdu Dataset

Instruction data for K–12 learning and teaching

Paper · Project Page · GitHub · Models · Quick Start · Citation

OmniEdu is a capability-oriented instruction-tuning dataset for educational models that connect subject knowledge, curriculum understanding, learner diagnosis, and instructional support. This repository releases 69,999 training examples, combining 60,951 education-specific examples with 9,048 general-purpose examples.

The dataset accompanies OmniEdu: Open Foundation Models for Learning and Teaching. The paper describes a mixture drawn from more than 100 educational resources and general instruction sources, containing 15.96M supervised response tokens.

Hao Liang · Qihan Lin · Meiyi Qiang · Linzhuang Sun · Hengyi Feng · Mingrui Chen · Sizhe Qiu · Wentao Zhang
Peking University · University of the Chinese Academy of Sciences · Zhongguancun Academy

Dataset at a glance

Property	Released data
Examples	69,999
Education-specific / general-purpose	60,951 / 9,048
Multimodal examples	12,146 (17.4%)
Languages	English and Chinese
Modalities	Text; image and text
Format	JSON Lines with ShareGPT-style messages
Configuration	core_v6_full_system_prompted
Split	train
System prompt IDs in use	20, including general_native
Image assets	Included under images/

This is the system-prompted Core V6 release. Every example starts with a system message. The assembly report records that non-system message content and image references are unchanged from the unprompted Core V6 assembly. This repository provides a training mixture; it does not define validation or test splits.

Capability composition

Capability family	Examples	Share	Training focus
Subject competence	31,855	45.51%	Solve school-level problems using textual and visual evidence.
Pedagogical action and scaffolding	14,226	20.32%	Provide hints, guiding questions, feedback, and adaptive instructional support.
Curriculum grounding	9,381	13.40%	Connect problems and concepts to knowledge points, curriculum standards, and prerequisite relations.
Diagnostic reasoning	5,489	7.84%	Identify learner errors, misconceptions, and gaps in understanding.
General-purpose instruction	9,048	12.93%	Maintain broad instruction-following behavior alongside educational specialization.
Total	69,999	100%

The four education-specific families account for 60,951 examples. General-purpose instruction supplies the remaining 9,048 examples. Percentages are rounded.

Quick start

Load the training split

pip install -U datasets huggingface_hub pillow

import json
from datasets import load_dataset

DATASET_ID = "lhpku20010120/Omni-Edu"
CONFIG = "core_v6_full_system_prompted"

train = load_dataset(DATASET_ID, CONFIG, split="train")
example = train[0]
metadata = json.loads(example["extra_info"])

print(len(train))  # 69999
print(example["messages"])
print(example["images"])
print(metadata["category"], metadata["system_prompt_id"])

For sequential inspection without first downloading the entire JSONL file, pass streaming=True to load_dataset and retrieve an example with next(iter(train)) instead of indexing.

Load an image from an example

The images field contains relative file paths, not decoded image objects. Loading the training split does not automatically download and open these image assets. Continue from the example above:

from huggingface_hub import hf_hub_download
from PIL import Image

loaded_images = []
for relative_path in example["images"]:
    local_path = hf_hub_download(
        repo_id=DATASET_ID,
        repo_type="dataset",
        filename=relative_path,
    )
    with Image.open(local_path) as image:
        loaded_images.append(image.convert("RGB"))

print([image.size for image in loaded_images])

Preserve the order of the image paths and any repeated entries: they correspond to the <image> placeholders in the conversation. For text-only examples, images is an empty list.

Download the full release for local training

This downloads the JSONL file, image assets, and accompanying metadata:

from pathlib import Path
from datasets import load_dataset
from huggingface_hub import snapshot_download

release_dir = Path(snapshot_download(
    repo_id="lhpku20010120/Omni-Edu",
    repo_type="dataset",
    local_dir="./Omni-Edu",
))
train = load_dataset(
    "json",
    data_files={"train": str(release_dir / "train.jsonl")},
    split="train",
)
image_paths = [release_dir / path for path in train[0]["images"]]
print(image_paths)

For reproducible experiments, set revision to a specific dataset commit when loading or downloading. LLaMA-Factory users can use the supplied dataset registration, which maps the message roles, content, and image fields.

Data format

Each line of train.jsonl is a JSON object with exactly three top-level fields:

Field	Type	Meaning
messages	List of objects	A system message followed by user and assistant turns; each object has role and content.
images	List of strings	Image paths relative to the repository root; an empty list for text-only examples.
extra_info	String containing JSON	Provenance, capability category, prompt ID, and optional selection or source metadata. Parse with json.loads.

The following is a schematic example; the text and image filename are placeholders, and the metadata is abbreviated:

{
  "messages": [
    {
      "role": "system",
      "content": "Task-specific instruction..."
    },
    {
      "role": "user",
      "content": "Problem:\n<image>Question text..."
    },
    {
      "role": "assistant",
      "content": "Target answer..."
    }
  ],
  "images": [
    "images/0a/example.png"
  ],
  "extra_info": "{\"category\": \"subject_competence\", \"dataset\": \"geometry3k-answer-only\", \"uid\": \"geometry3k-513\", \"system_prompt_id\": \"solve_answer_only\"}"
}

Conversations contain up to 21 messages; approximately 94.5% are single-turn. Examples contain zero to 16 image references: approximately 82.6% are text-only, 16.2% contain one image, and the remainder contain multiple images. In 131 rows, an image is referenced more than once.

After parsing extra_info, common keys include category, dataset, uid, system_prompt_id, and system_prompt_version. Optional fields include kcenter, source_license, and language; availability varies by source.

System prompts

Task-specific instructions make the expected teaching or problem-solving behavior explicit. Examples of active prompt IDs include:

Prompt ID	Intended behavior
solve_answer_only	Return an answer in the requested format without an explanation.
solve_reasoned	Produce a self-contained solution with a clear final answer.
knowledge_point	Identify concepts, standards, prerequisites, or curriculum relations.
diagnose_and_correct	Identify the learner's first material error and provide a useful correction.
mathtutor_scaffolding	Give a concise next scaffold.
multiturn_socratic	Use dialogue history to guide the learner with focused questions.
general_native	Retain the upstream system prompt for general-purpose examples.

There are 20 prompt IDs in use in the assembly report. Education-specific instructions are defined in system_prompt_registry.json. The 9,048 general_native examples retain their upstream prompts, so general_native is not an entry in that registry. The registry also contains general_assistant, which is not used in this release's prompt counts.

Construction and provenance

The paper describes a pipeline combining deterministic cleaning, semantic auditing and rewriting, task-specific quality scoring, token-budgeted diversity selection, and pedagogical instruction assignment. Per-example metadata identifies the contributing source and, where provided upstream, its license.

This release was assembled as training_core_v6/releases/v6_full_system_prompted. Its assembly report records:

Assembly check	Result
Every row has a system message	Yes
Non-system content unchanged from the unprompted assembly	Yes
Image references unchanged	Yes
Tokenizer used for length audit	Qwen/Qwen3.5-9B-Base
Maximum audited sequence length	12,786 tokens
99th-percentile sequence length	8,830 tokens
Audit cutoff	16,384 tokens
Examples above the cutoff	0

These are dataset assembly statistics under the named tokenizer. The paper separately reports a 32,768-token maximum training sequence length for model fine-tuning; it is a training configuration rather than the longest sequence in this release. The 15.96M supervised response tokens reported in the paper measure response-token volume, not total input-plus-output sequence length.

Repository contents

Path	Contents
train.jsonl	Training examples
images/	Referenced image assets
system_prompt_registry.json	Task-specific instruction definitions
assembly_report.json	Prompt counts and assembly audit
dataset_info.json	LLaMA-Factory dataset registration

Model family

Explore the accompanying models for inference examples and evaluation results:

Model	Scale
OmniEdu-4B	4B
OmniEdu-9B	9B
OmniEdu-27B	27B

Intended use and limitations

The mixture supports supervised fine-tuning, analysis of educational instruction data, and research on curriculum grounding, learner diagnosis, and tutoring. Coverage reflects the contributing sources and does not establish uniform quality across subjects, languages, curricula, or learner populations. Inspect examples for suitability before using them in educational applications. Model benchmark results reported in the paper do not establish classroom learning gains.

Source licenses and usage terms

This repository retains the other license designation. Licenses and usage terms of the constituent datasets and source materials continue to apply. Check extra_info.dataset to identify the source and extra_info.source_license where available; consult the upstream terms before reuse or redistribution. Missing per-row license metadata does not imply unrestricted permission.

Citation

If you use the dataset or accompanying models in your research, please cite:

@misc{liang2026omniedu,
  title         = {OmniEdu: Open Foundation Models for Learning and Teaching},
  author        = {Hao Liang and Qihan Lin and Meiyi Qiang and Linzhuang Sun and Hengyi Feng and Mingrui Chen and Sizhe Qiu and Wentao Zhang},
  year          = {2026},
  eprint        = {2609.23088},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2609.23088}
}

Feedback

For questions about the data, source attribution, or educational use cases, open a discussion in the dataset Community tab or an issue on GitHub.