lfaviate/China-K12-STEM-10K-CoT-Reasoning

K12-STEM-CoT-Chinese

1.54M Chinese K12 STEM problems with chain-of-thought solutions, 48% with diagrams. The largest structured Chinese math/physics/chemistry reasoning dataset.

This is a curated sample (10,000 problems) of the full 1.54M dataset available via API.

Full Dataset Access

Access the full 1,540,000+ problems via API →

	This Sample	Full API
Total problems	10,025	1,540,000+
With CoT solutions	10,025	1,490,000+
With diagrams	6,093	740,000+
Subjects	Math / Physics / Chemistry	Same
Knowledge points	Partial	4,470 nodes (full tree)
Filtering & Search	No	By grade / topic / difficulty / type / year
Full-text search	No	Chinese + English keywords
Updated	Static snapshot	Continuously
Price	Free	Free tier available

Dataset Description

10,025 Chinese K12 (Grades 7–12) STEM exam questions with human-written step-by-step solutions. Covers Mathematics, Physics, and Chemistry from real Chinese school exams.

Key Features

• LaTeX-formatted content with proper math notation ($x^{2}$, $CO_{2}$)
• Chain-of-Thought solutions — average 11.2 steps per problem, human-written (not LLM-generated)
• Multimodal — 60.8% of problems include diagrams (physics circuits, geometry figures, chemistry apparatus)
• Bilingual knowledge points — Chinese + English labels for every problem
• Difficulty calibrated — 4-level difficulty scoring (Easy / Medium / Medium-Hard / Hard)
• Diverse question types — Multiple-choice, fill-in-the-blank, calculation, proof, experiment, and more

Statistics

Metric	Value
Total problems	10,025
Mathematics	4,006
Physics	3,005
Chemistry	3,014
Grades	7–12
Difficulty	Easy 20% / Medium 30% / Medium-Hard 45% / Hard 5%
Avg CoT steps	11.2
With diagrams	60.8%

Split	Count
train	9,023
test	501
validation	501

Grade Distribution

Grade	Count
7	351
8	977
9	3,554
10	1,333
11	1,384
12	2,426

Use Cases

1. AI Reasoning Training — Fine-tune math/science reasoning models (Qwen-Math, DeepSeek-Math, etc.) with real human-written CoT data
2. Multimodal STEM — Train vision-language models on diagram-based physics/chemistry problems
3. LLM Evaluation — Benchmark Chinese STEM reasoning ability across difficulty levels
4. Educational AI — Build adaptive tutoring systems with knowledge-point-aware question selection
5. Curriculum Analytics — Analyze knowledge-point coverage and difficulty distribution across grades

Data Fields

Field	Type	Description
question_id	string	Unique identifier
subject	string	Mathematics / Physics / Chemistry
grade	int	7–12
grade_level	string	e.g., "Grade 9"
difficulty	string	Easy / Medium / Medium-Hard / Hard
difficulty_score	float	0–1 (lower = harder)
question_type	string	Multiple-choice, Calculation, Experiment, etc.
question_text	string	LaTeX-formatted question text
question_images	list[str]	CDN URLs for diagrams (JSON-encoded in Parquet)
options	list[str]	Answer options for multiple-choice (JSON-encoded in Parquet)
answer	string	Final answer
method_analysis	string	Solution methodology explanation
step_by_step_proof	string	Detailed proof/solution (HTML format)
solution_steps	list[str]	Individual CoT reasoning steps (JSON-encoded in Parquet)
step_count	int	Number of solution steps
knowledge_points	list[str]	Knowledge point labels in Chinese (JSON-encoded in Parquet)
knowledge_points_en	list[str]	Knowledge point labels in English (JSON-encoded in Parquet)
has_image	bool	Whether the problem includes diagrams
onshelf_date	string	Date the problem was published

Example

{
  "question_id": "b45f681d7de247e5...",
  "subject": "Physics",
  "grade": 9,
  "difficulty": "Medium-Hard",
  "difficulty_score": 0.4,
  "question_type": "Comprehensive",
  "question_text": "固定不动的救援车将陷入沙滩中的小轿车沿水平直线方向拉出。若小轿车受到的重力为12000N...",
  "question_images": ["https://assets.stem-science.com/images/e3/e332189096407164.png"],
  "options": [],
  "answer": "（1）①5m；②1800W；③80%",
  "solution_steps": [
    "小轿车被匀速拉动通过的路程是 $s_1 = vt = 0.5\\text{m/s} \\times 10\\text{s} = 5\\text{m}$",
    "拉力 $F_1$ 做的功为 $W_{有} = F_1 s_1 = 3.6 \\times 10^3\\text{N} \\times 5\\text{m} = 1.8 \\times 10^4\\text{J}$",
    "功率为 $P = \\frac{W_{有}}{t} = \\frac{1.8 \\times 10^4\\text{J}}{10\\text{s}} = 1800\\text{W}$",
    "机械效率为 $\\eta = \\frac{W_{有}}{W_{总}} = \\frac{1.8 \\times 10^4}{2 \\times 10^4} = 80\\%$"
  ],
  "step_count": 13,
  "knowledge_points": ["简单机械综合", "功率的计算"],
  "knowledge_points_en": ["Simple Machine Synthesis", "Power Calculation"],
  "has_image": true
}

Quick Start

from datasets import load_dataset

# Load dataset
ds = load_dataset("lfaviate/China-K12-STEM-10K-CoT-Reasoning")

# Filter by subject
physics = ds["train"].filter(lambda x: x["subject"] == "Physics")

# Filter multimodal problems (with diagrams)
multimodal = ds["train"].filter(lambda x: x["has_image"] == True)

# Stream for memory efficiency
ds_stream = load_dataset("lfaviate/China-K12-STEM-10K-CoT-Reasoning", split="train", streaming=True)

Comparison with Other Datasets

Dataset	Size	Subjects	CoT	Multimodal	Language
K12-STEM-CoT-Chinese (this)	10K (1.54M full)	Math+Phys+Chem	97%	61%	zh
GSM8K	8.5K	Math	✓	✗	en
MATH	12.5K	Math	✓	✗	en
NuminaMath-CoT	~900K	Math	✓	✗	zh+en
GAOKAO-Bench	~2K	Multi	✗	Partial	zh
CMM-Math	28K	Math	✓	✓	zh

Unique advantages: (1) Covers Physics + Chemistry, not just Math; (2) 61% multimodal; (3) Human-written CoT, not LLM-generated; (4) 1.54M full scale via API.

Limitations

This sample contains 10,025 problems — 0.65% of the full dataset. For production use cases requiring:

• Full scale: 1,540,000+ problems across all 4,470 knowledge points
• Advanced filtering: By grade, topic, difficulty, type, year, knowledge point
• Full-text search: Chinese and English keyword search
• Knowledge-point tree: 4,470-node hierarchical navigation
• Regular updates: New exam questions added continuously
• High throughput: Up to 50 items/request, 50 req/sec

Use the full API on RapidAPI →

Citation

@dataset{k12_stem_cot_chinese_2026,
  title={K12-STEM-CoT-Chinese: A Large-Scale Chinese K12 STEM Dataset with Chain-of-Thought Solutions},
  year={2026},
  url={https://huggingface.co/datasets/lfaviate/China-K12-STEM-10K-CoT-Reasoning},
  note={Sample of 10,025 problems from 1.54M total. Full dataset available via API.}
}

License

CC BY-NC-SA 4.0

• BY — Attribution required
• NC — Non-commercial use only
• SA — Derivative works must use the same license

For commercial licensing or bulk data access (500K+ problems), please contact us.