TextbooksToRL

A tool that automatically processes textbooks to generate structured question-answer datasets for machine learning and reinforcement learning training.

How It Works

1. Extract Text: Converts PDF textbooks to text format
2. Process Pages: Splits content into manageable chunks (3-5 pages)
3. Generate Questions: Uses AI models to create questions from each chunk
4. Verify Solutions: Checks answers by feeding page back in with question to ensure same answer
5. Output Dataset: Saves structured JSON files with questions, solutions, and metadata

The tool supports multiple difficulty levels (high school → PhD) and various AI models (OpenAI, Anthropic, DeepSeek, etc.).

Setup

1. Install dependencies:

uv sync

2. Set up API keys:

export OPENAI_API_KEY='your-key'
export ANTHROPIC_API_KEY='your-key'
export DEEPSEEK_API_KEY='your-key'
export DEEPINFRA_API_KEY='your-key'

Usage

Basic Usage

# Generate questions from textbooks
uv run textbooks-to-rl --help

# Run with custom settings
uv run textbooks-to-rl \
  --model "Qwen/QwQ-32B" \
  --output-dir "my_questions" \
  --difficulty undergrad \
  --verbose

Step 1: Add Textbooks

Place PDF files in textbooks/pdfs/ or process them:

uv run python scripts/process_pdfs.py

Step 2: Generate Questions

# Basic generation
uv run textbooks-to-rl

# With custom options
uv run textbooks-to-rl \
  --pages-per-group 5 \
  --batch-size 50 \
  --questions-per-chunk 8 \
  --difficulty grad \
  --no-verify

Step 3: Filter Questions (Optional)

uv run python scripts/filter.py \
  --folders generated_questions \
  --output-dir filtered_results \
  --model gpt-4o-mini

Options

Option	Description	Default
--model	AI model for generation	Qwen/QwQ-32B
--output-dir	Output directory	generated_questions
--textbooks-dir	Textbooks directory	textbooks/txt
--pages-per-group	Pages per processing group	3
--batch-size	Parallel batch size	100
--questions-per-chunk	Questions per chunk	10
--difficulty	Question difficulty level	undergrad
--no-verify	Skip solution verification	False
--verbose	Enable debug logging	False

Development

# Install with dev dependencies
uv sync --dev

# Run code quality checks
uv run ruff check src/ --fix
uv run pytest