Chapter 4 Released: Data-Driven Depth Pruning with Cosine Similarity

[peremartra]

🎉 Just shipped Chapter 4!

This one dives deep into intelligent data-driven layer selection using PyTorch hooks + cosine similarity analysis.

Key takeaways:

• How to measure which layers actually contribute to your task
• Why layer importance varies drastically between simple vs complex text
• Building specialized models (11-13% speedup) optimized for specific data types

🔗 Notebook: https://github.com/peremartra/Rearchitecting-LLMs/blob/main/CH04/CH04_NB01_Cosine_Similarity.ipynb
📚 Also updated the main README with a complete learning path (CH2→CH3→CH4)

Questions? Drop them here. I'm actively working on CH5 (Width Pruning) and would love to hear what you'd like to see next.

---

For those using optiPfair in production: This chapter shows the research behind analyze_layer_importance() and how to use it for real-world pruning decisions.