tanmay-kalbande/pustakam-edu-corpus

Pustakam Educational Corpus

A synthetic-data testbed for studying how different LLMs write, structure, and teach — 1,000+ AI-generated books built for model comparison, readability research, and curriculum-design analysis, not for classroom use.

A large-scale structured synthetic educational text corpus generated by an automated multi-model curriculum pipeline. Contains 1,000+ books totaling 23+ million words across 70 categories and 3 difficulty levels, with rich metadata including model provenance, text quality metrics, and structural annotations.

Dataset Description

Summary

Pustakam is a corpus of AI-generated educational books produced by an automated pipeline that routes content generation across multiple LLM providers (Z.ai GLM, Cerebras, Mistral). Each book follows a consistent structure: AI-generated roadmap → sequential chapter generation with cross-chapter continuity → introduction, summary, and glossary.

This is a synthetic dataset. The content was generated by AI models and has not been verified for factual accuracy. It is intended for research on AI content generation patterns, model comparison, educational text structure, and readability analysis — not as a factual knowledge source.

Key Properties

Property	Value
Total Books	~1,091
Total Words	~23.7 million
Total Chapters	~10,892
Categories	70
Difficulty Levels	beginner, intermediate, advanced
Editions	stellar (professional), street (raw/informal), desi (Hinglish)
Models	8 distinct (GLM-5.2, GPT-OSS-120B, Gemma-4-31B, etc.)
Language	English (primary), Hinglish (23 books)
Generation Period	July–August 2026

What Makes This Dataset Unique

1. Multi-model provenance: Every book records which model(s) generated it. Same prompts → different models → comparable outputs. This enables controlled model comparison research.
2. Structured educational format: Not free-form text. Each book has a roadmap, chapters with section headings, introduction, summary, and glossary. Schema is consistent across all books.
3. Cross-chapter continuity: The pipeline injects previously-defined terms into subsequent chapter prompts, creating genuine cross-referencing between chapters (not just independent essays).
4. Quality-gated: Books that failed structural quality checks (too few chapters, thin content, placeholder text, duplicate titles) were rejected before entering the corpus.
5. Text quality metrics: Each record includes computed readability scores (Flesch-Kincaid), vocabulary diversity (Type-Token Ratio), sentence statistics, and structural annotations.

Dataset Configurations

book_metadata (Lightweight)

One row per book. Metadata only — no content text. Best for exploratory analysis, filtering, and understanding corpus distribution.

Column	Type	Description
slug	string	URL-safe identifier
title	string	Book title
goal	string	Learning objective
category	string	Topic category (70 unique)
tags	list[string]	Topic tags
language	string	Language code (en, hi)
complexity	string	beginner / intermediate / advanced
edition	string	stellar / street / desi
word_count	int	Total word count
chapter_count	int	Number of chapters
reading_time_mins	int	Estimated reading time
meta_description	string	SEO meta description
model_used	string	Model(s) that generated the book
generated_at	string	ISO 8601 timestamp

chapters (Chapter-Level)

One row per chapter. Includes full chapter text and per-chapter quality metrics. Best for text analysis, readability research, and model comparison at the chapter level.

Column	Type	Description
book_slug	string	Parent book identifier
book_title	string	Parent book title
chapter_index	int	0-indexed position in book
chapter_title	string	Chapter title
content	string	Full chapter text (Markdown)
word_count	int	Chapter word count
category	string	Topic category
complexity	string	Difficulty level
edition	string	Edition persona
model_used	string	Model(s) used
position_in_book	string	first / early / middle / late / last
total_chapters	int	Total chapters in parent book
has_code_blocks	bool	Contains code blocks
code_block_count	int	Number of code blocks
heading_count	int	Number of section headings
type_token_ratio	float	Vocabulary diversity (0-1)
fk_grade_level	float	Flesch-Kincaid grade level
avg_sentence_length	float	Average words per sentence

books_full (Full Text)

One row per book. Includes the complete assembled book text (introduction + chapters + summary + glossary) and book-level quality metrics. Best for full-document analysis and NLP tasks.

Column	Type	Description
slug	string	URL-safe identifier
title	string	Book title
goal	string	Learning objective
category	string	Topic category
tags	list[string]	Topic tags
language	string	Language code
complexity	string	Difficulty level
edition	string	Edition persona
word_count	int	Total word count
chapter_count	int	Number of chapters
reading_time_mins	int	Estimated reading time
model_used	string	Model(s) used
generated_at	string	ISO 8601 timestamp
full_text	string	Complete book in Markdown
type_token_ratio	float	Vocabulary diversity
fk_grade_level	float	Flesch-Kincaid grade level
avg_sentence_length	float	Average words per sentence
code_block_count	int	Number of code blocks
has_introduction	bool	Contains introduction section
has_summary	bool	Contains summary section
has_glossary	bool	Contains glossary section

Usage

from datasets import load_dataset

# Load metadata only (fast)
metadata = load_dataset("tanmay-kalbande/pustakam-edu-corpus", "book_metadata")

# Load all chapters with content
chapters = load_dataset("tanmay-kalbande/pustakam-edu-corpus", "chapters")

# Load full books
books = load_dataset("tanmay-kalbande/pustakam-edu-corpus", "books_full")

# Example: Filter programming books at advanced level
advanced_prog = chapters["train"].filter(
    lambda x: x["category"] == "programming" and x["complexity"] == "advanced"
)

# Example: Compare word counts across models
import pandas as pd
df = metadata["train"].to_pandas()
print(df.groupby("model_used")["word_count"].describe())

Research Use Cases

Recommended Uses

• Model comparison: Compare output quality (length, readability, vocabulary, structure) across 8 models using the same prompt pipeline
• Educational text structure analysis: Study heading density, section patterns, introduction/summary presence at scale
• Readability research: Analyze whether labeled difficulty (beginner/intermediate/advanced) correlates with actual readability scores
• Synthetic text detection: Train classifiers on a large corpus of known-synthetic educational text
• Curriculum design patterns: Study how AI structures learning progressions across chapter sequences
• Code-switching linguistics: The Hinglish (desi) edition provides structured Hindi-English code-switching in educational context

Not Recommended

• Fine-tuning for factual knowledge: Content is AI-generated and not verified for accuracy. May contain hallucinations.
• Medical/legal/financial advice: Generated content in health, finance, and career categories has not been reviewed by domain experts.
• Ground truth for any domain: Do not treat this as authoritative educational material.

Known Limitations & Roadmap

#	Limitation	Detail	Planned Fix
1	No factual accuracy verification	All content is AI-generated. No domain expert review has been performed.	Out of scope for this dataset — flagged clearly as a "not for factual use" corpus rather than fixed.
2	Readability-difficulty mismatch	The average Flesch-Kincaid grade level is 3.4 across all complexity levels. Even "advanced" books use simple sentence structures.	Investigating stronger complexity-conditioning in the generation prompts for a future release.
3	Beginner-heavy distribution	67% of books are beginner-level, 24% intermediate, 9% advanced.	Rebalancing planned by adjusting topic-seed sampling toward intermediate/advanced goals.
4	Category imbalance	Business has 99 books; median category has 10.	Quota caps per category under consideration for the next generation run.
5	8 near-duplicate titles	The deduplication system (Jaccard keyword similarity) missed some semantically similar topics.	Evaluating embedding-based similarity checks to replace/augment the keyword-based dedup.
6	Model distribution skew	GLM-5.2 generated 40% of books; smaller models are underrepresented.	Planned quota rebalancing across all 8 models in the next generation batch (see note under Models Used).

Quality Assessment

Based on analysis of 20 stratified sample books:

Metric	Mean	Interpretation
Type-Token Ratio	0.116	Low-medium vocabulary diversity (typical for educational text)
Avg Sentence Length	15.0 words	Good readability
Flesch-Kincaid Grade	3.4	Very accessible (4th grade level)
Heading Density	9.6 per 1K words	Well-structured sections
Chapter Word Count CV	0.145	Consistent chapter lengths
Structural Completeness	90% intro, 90% summary, 80% glossary	Strong structural consistency

Dataset Creation

Source Data

Generated by the Pustakam automated curriculum pipeline. The pipeline takes topic seeds (learning goals with category, tags, and difficulty), generates a structured roadmap, then produces sequential chapters with cross-chapter continuity via glossary-term injection.

Generation Pipeline

1. Topic seeding: AI-generated or human-curated learning goals
2. Roadmap generation: AI produces a chapter-by-chapter curriculum plan (6-14 chapters)
3. Sequential chapter generation: Each chapter prompt includes the full roadmap, previous chapter context, and extracted key terms
4. Assembly: AI generates an introduction, summary, and glossary for each book
5. Quality gate: Books failing structural checks (min chapters, min word count, no placeholders, no duplicate titles) are rejected

Models Used

Model	Provider	Books Generated	Avg Word Count
glm-5.2	Z.ai	~441 (40%)	23,475
mistral-small-2506	Mistral	~232 (21%)	23,170
gpt-oss-120b	Cerebras	~224 (21%)	21,420
gemma-4-31b	Cerebras	~160 (15%)	14,375
mistral-large-latest	Mistral	~22 (2%)	28,692
Others	Various	~12 (<1%)	Varies

Note: mistral-large-latest and the "Others" group are underrepresented mainly due to tighter rate limits during the generation run, not a deliberate quality choice. A future release aims to rebalance generation quotas more evenly across all 8 models.

Citation

@dataset{kalbande2026pustakam,
  title={Pustakam Educational Corpus: A Large-Scale Structured Synthetic Educational Text Dataset},
  author={Kalbande, Tanmay},
  year={2026},
  publisher={HuggingFace},
  url={https://huggingface.co/datasets/tanmay-kalbande/pustakam-edu-corpus}
}

License

This dataset is released under the Creative Commons Attribution 4.0 International License (CC-BY-4.0).

Contact

• Author: Tanmay Kalbande
• Website: tanmaysk.in
• GitHub: tanmay-kalbande
• Library: tanmaysk.in/library