Datasets:
language:
- en
license: apache-2.0
size_categories:
- 10M<n<100M
task_categories:
- text-generation
tags:
- pretraining
- educational
- pedagogical
- synthetic
- sutra
- multi-domain
- 10B
pretty_name: Sutra 10B Pretraining Dataset
Sutra 10B Pretraining Dataset
A high-quality pedagogical dataset designed for LLM pretraining, containing 10,193,029 educational entries totaling over 10 billion tokens. This is the largest dataset in the Sutra series, designed to demonstrate that dense, curated datasets can provide best-in-class pretraining performance for small language models.
Dataset Description
This dataset was generated using the Sutra framework, which creates structured educational content optimized for language model pretraining. Each entry is designed to maximize learning efficiency through:
- Clear pedagogical structure: Content follows proven educational patterns
- Cross-domain connections: Concepts are linked across disciplines
- Varied complexity levels: From foundational (level 1) to advanced (level 10)
- Quality-controlled generation: All entries meet minimum quality thresholds
- Diverse content types: 33 different pedagogical formats
- Rich metadata: Every entry annotated with 13 structured fields
Dataset Statistics
| Metric | Value |
|---|---|
| Total Entries | 10,193,029 |
| Total Tokens | 10,218,677,925 |
| Avg Tokens/Entry | 1002 |
| Avg Quality Score | 0.701 |
| Tokenizer | SmolLM2 (HuggingFaceTB/SmolLM2-135M) |
Domain Distribution
| Domain | Entries | Tokens | Percentage |
|---|---|---|---|
| interdisciplinary | 3,561,052 | 3570.0M | 34.9% |
| technology | 2,154,481 | 2159.9M | 21.1% |
| science | 1,456,708 | 1460.3M | 14.3% |
| social_studies | 862,288 | 864.4M | 8.5% |
| mathematics | 830,414 | 832.5M | 8.1% |
| life_skills | 559,667 | 561.1M | 5.5% |
| arts_and_creativity | 455,738 | 456.9M | 4.5% |
| language_arts | 235,957 | 236.5M | 2.3% |
| philosophy_and_ethics | 76,724 | 76.9M | 0.8% |
Content Type Distribution (Top 15)
| Content Type | Count | Percentage |
|---|---|---|
| historical_context | 3,082,957 | 30.2% |
| concept_introduction | 928,244 | 9.1% |
| data_analysis | 776,495 | 7.6% |
| worked_examples | 697,861 | 6.8% |
| problem_set | 676,977 | 6.6% |
| tutorial | 620,163 | 6.1% |
| technical_documentation | 520,246 | 5.1% |
| research_summary | 494,023 | 4.8% |
| code_implementation | 473,056 | 4.6% |
| practical_application | 438,157 | 4.3% |
| creative_writing | 337,065 | 3.3% |
| reasoning_demonstration | 227,343 | 2.2% |
| qa_pairs | 200,076 | 2.0% |
| ethical_analysis | 157,882 | 1.5% |
| experiment_design | 141,859 | 1.4% |
Data Sources
Sutra-10B was created by scaling the same recipe used for Sutra-1B from 1 billion to 10 billion tokens. The core pedagogical content was generated using the Sutra framework, then mixed with several high-quality open datasets for diversity:
| Source | Description | Approximate Tokens |
|---|---|---|
| Sutra (core) | Pedagogical content generated with the Sutra framework, scaled from the 1B recipe | ~7.8B |
| Nemotron-CC-Math v1 | High-quality mathematical content (NVIDIA) | ~0.5B |
| OpenWebMath | Mathematical web content | ~0.5B |
| Wikipedia (English) | Encyclopedic knowledge | ~0.5B |
| Cosmopedia | Synthetic educational content (multiple subsets) | ~0.5B |
| FineWeb-Edu | High-quality educational web content | ~0.5B |
Data Fields
Each entry contains 13 structured fields:
| Field | Type | Description |
|---|---|---|
id |
string | Unique identifier (UUID) |
concept_name |
string | The concept being taught (2-5 words) |
domain |
string | Primary knowledge domain (9 domains) |
content_type |
string | Type of pedagogical content (33 types) |
text |
string | The main educational content |
quality_score |
float | Quality assessment score (0.0-1.0) |
information_density |
string | Measure of information per token (low/medium/high) |
complexity_level |
integer | Difficulty level (1-10) |
token_count |
integer | Number of tokens (SmolLM2 tokenizer) |
prerequisites |
list[string] | Required prior knowledge concepts |
builds_to |
list[string] | Advanced concepts this enables |
cross_domain_connections |
list[string] | Related knowledge domains |
quality_assessment |
object | Multi-dimensional quality scores |
Quality Assessment Sub-fields
| Sub-field | Type | Description |
|---|---|---|
clarity |
float | How clear and readable (0.0-1.0) |
accuracy |
float | Factual correctness (0.0-1.0) |
pedagogy |
float | Educational structure quality (0.0-1.0) |
engagement |
float | How engaging the content is (0.0-1.0) |
depth |
float | Depth of coverage (0.0-1.0) |
creativity |
float | Creative presentation (0.0-1.0) |
Valid Domains (9)
mathematics, science, technology, language_arts, social_studies, arts_and_creativity, life_skills, philosophy_and_ethics, interdisciplinary
Valid Content Types (33)
concept_introduction, reasoning_demonstration, code_implementation, technical_documentation, tutorial, cross_domain_bridge, worked_examples, qa_pairs, common_misconceptions, meta_learning, synthesis, prerequisite_scaffolding, code_explanation, diagnostic_assessment, code_debugging, historical_context, research_summary, problem_set, case_study, analogy, experiment_design, proof, algorithm_analysis, data_analysis, ethical_analysis, comparative_analysis, creative_writing, debate_argument, practical_application, thought_experiment, visualization, system_design, review_summary
Data Cleaning
The dataset underwent comprehensive cleaning:
- Deduplication: SHA-256 hash-based exact duplicate removal across all sources
- Quality Filtering: Entries below quality_score 0.3 removed
- Length Filtering: Entries shorter than 50 tokens or longer than 65,536 tokens removed
- Garbage Detection: Repetitive content, control characters, non-English content filtered
- Field Validation: All 13 fields validated and normalized
Metadata Generation
Metadata was generated using heuristic keyword-based classification:
- Domain and content type classification via pattern matching and text analysis
- Quality scores computed from text statistics (vocabulary diversity, structure, length)
- Token counts computed using SmolLM2 tokenizer for accuracy
Usage
from datasets import load_dataset
# Load the full dataset
ds = load_dataset("codelion/sutra-10B", split="train")
# Stream for large-scale training
ds = load_dataset("codelion/sutra-10B", split="train", streaming=True)
# Filter by domain
math_ds = ds.filter(lambda x: x["domain"] == "mathematics")
# Filter by quality
high_quality = ds.filter(lambda x: x["quality_score"] > 0.7)
# Filter by complexity
beginner = ds.filter(lambda x: x["complexity_level"] <= 3)
Scaling Trajectory
Sutra-10B is the largest dataset in the Sutra series, scaling the original 1B recipe by 10x. When evaluated on SmolLM2-70M (69M parameters), benchmark performance remains consistent across scales, suggesting the model has reached its capacity ceiling. Larger models are expected to benefit more from the additional data and diversity.
Intended Use
This dataset is designed for:
- LLM Pretraining: High-quality educational content for foundational model training
- Domain-specific fine-tuning: Subset by domain for specialized training
- Educational AI research: Studying pedagogical content generation
- Curriculum learning: Progressive complexity for staged training
- Small model optimization: Demonstrating data quality > quantity for small LMs
Related Datasets
- sutra-1B: 1B token pretraining dataset
- sutra-100M: 100M token subset
- sutra-30k-seeds: Instruction prompts for post-training
- sutra-magpie-sft: SFT dataset
Citation
@article{sharma2026sutra,
title={Scaling Pedagogical Pretraining: From Optimal Mixing to 10 Billion Tokens},
author={Sharma, Asankhaya},
year={2026},
url={https://huggingface.co/blog/codelion/scaling-pedagogical-pretraining-10-billion-tokens}
}
License
Apache 2.0