Recommendations for Using This Script

Model Selection

• DeepSeek-R1 (default) - Excellent for complex literary translation requiring deep reasoning about style and nuance
• Qwen3 - Good alternative with thinking capabilities, faster for simpler texts
• Consider the model size based on your hardware:
  • 8B models for consumer GPUs (RTX 3060+)
  • 32B+ models for professional setups with more VRAM

Optimal Configuration Settings

1. Temperature (0.7 default)

   • 0.5-0.6 for technical or historical texts requiring accuracy
   • 0.7-0.8 for literary fiction (good balance)
   • 0.8-0.9 for poetry or highly stylistic prose

2. Page Size

   • 300-500 words for dense philosophical texts (Kant, Hegel)
   • 500-800 words for narrative fiction
   • 200-400 words for poetry or highly stylized prose

3. Author Style Matching

   • Knausgård - Best for diary-like texts, memoirs, autobiographical works
   • Cusk - Ideal for philosophical dialogues, observational narratives
   • Ernaux - Perfect for sociological texts, class-conscious narratives
   • Sebald - Excellent for historical fiction, travelogues with melancholic tone

Text Preparation Tips

1. Pre-process Gutenberg files:

   # Remove Gutenberg headers/footers
   sed -n '/\*\*\* START OF THIS PROJECT/,/\*\*\* END OF THIS PROJECT/p' input.txt > clean.txt

2. For best results, convert HTML/EPUB to plain text first:

   pandoc input.epub -t plain -o input.txt

Performance Optimization

1. GPU Memory Management:

   • Close other applications using VRAM
   • Use smaller model variants if encountering OOM errors
   • Increase DELAY_BETWEEN_PAGES to prevent thermal throttling

2. Processing Strategy:

   • Run in test mode first (--test --pages 10) to verify style quality
   • Process challenging sections separately with adjusted prompts
   • Use thinking logs to debug translation issues

Quality Assurance

1. Review thinking logs - They reveal the model's reasoning about style choices
2. Compare multiple runs with different temperatures for critical passages
3. Post-process for consistency in:
   • Character names (maintain original or fully translate)
   • Place names and cultural references
   • Formatting of dialogue and internal monologue

Recommended Workflow

1. Initial Test (5-10 pages):

   python translator.py source.txt --test --pages 10 --author "Rachel Cusk"

2. Adjust parameters based on test output quality

3. Full translation with optimized settings:

   python translator.py source.txt --model deepseek-r1 --author "Karl Ove Knausgård"

4. Post-processing:

   • Review for consistency
   • Merge chapter breaks if needed
   • Final editing pass for flow

Advanced Usage

• Extend author styles by adding entries to AUTHOR_STYLES dictionary
• Chain multiple models: Use thinking model for complex passages, faster model for simple narrative
• Implement glossary: Add consistent translation rules for recurring terms
• Version control: Save different temperature/style attempts for comparison

Common Pitfalls to Avoid

1. Don't process texts with heavy formatting (footnotes, tables) without pre-processing
2. Avoid very large page sizes (>1000 words) - quality degrades
3. Don't skip the thinking mode - it significantly improves style consistency
4. Remember to check copyright status - ensure texts are truly public domain

The script is designed to be both powerful and flexible. Start with test runs, iterate on settings, and gradually refine based on the specific text characteristics and your quality requirements.