Skip to content

cli_pdf_qa.md

Latest commit

 

History

History
85 lines (68 loc) · 2.95 KB

cli_pdf_qa.md

File metadata and controls

85 lines (68 loc) · 2.95 KB
Raw

Q&A Generation Engine (cli_pdf_qa.py)

Purpose

Extracts Q&A pairs from PDF documents using GPU-accelerated language models with configurable difficulty and creativity parameters.

How It Works

Text Processing Pipeline

  1. PDF Extraction: PyMuPDF extracts raw text while preserving structure
  2. Chunking: Splits text into overlapping windows (default: 800 words, 100-word overlap)
  3. Chunk Scoring: Filters chunks based on information density and readability

Q&A Generation Strategy

Uses prompt templates to generate questions at different complexity levels:

  • Basic: Direct factual questions about explicit information
  • Intermediate: Analytical questions requiring understanding of relationships
  • Advanced: Synthesis questions combining multiple concepts

Creativity Control via Sampling

  • Conservative (T=0.3): Focused, literal questions with low variation
  • Balanced (T=0.7): Standard technical questions with moderate creativity
  • High Creativity (T=0.9): Exploratory questions with diverse phrasings

GPU Optimization

  • Model Loading: Llama-3-8B-Instruct with FP16 precision
  • Batch Processing: Processes multiple chunks simultaneously
  • Memory Management: Optional 8-bit quantization for 24GB+ models

Technical Implementation

Core Components

class PDFQAExtractor:
    - load_model()      # Initialize Llama-3 with GPU acceleration
    - extract_text()    # PDF → structured text chunks
    - generate_qa()     # Chunk → Q&A pairs via LLM
    - filter_quality()  # Remove low-quality outputs

Prompt Engineering

  • Domain-specific vocabulary injection
  • Context-aware question generation
  • Answer grounding to source text
  • Hallucination prevention through strict prompting

Configuration

Key Parameters

  • --difficulty-levels: basic, intermediate, advanced
  • --temperature: Sampling randomness (0.1-1.0)
  • --chunk-size: Text window size in words
  • --batch-size: GPU memory vs speed tradeoff
  • --quantize: Enable 8-bit model compression

Output Format

{
  "instruction": "What is the purpose of the preamp gain control?",
  "input": "",
  "output": "The preamp gain control adjusts input sensitivity...",
  "source_chunk": "Original PDF text segment",
  "difficulty": "basic",
  "chunk_id": "chunk_042",
  "quality_score": 0.87
}

Performance Characteristics

Typical Throughput

  • L40S GPU: ~50 Q&A pairs/minute with quantization
  • Memory Usage: 12-16GB VRAM (8-bit), 20-24GB (FP16)
  • Quality vs Speed: Higher temperatures = more creativity but slower generation

Expected Outputs

  • Per Difficulty Level: ~50-100 pairs from typical manual
  • Total (3×3 Matrix): 500-1000 pairs across all combinations
  • Quality Distribution: 70-80% meet basic quality thresholds

Use Cases

  • Technical documentation → training data
  • Domain-specific Q&A dataset creation
  • Knowledge extraction from manuals
  • Multi-level educational content generation