Search is not available for this dataset
image image |
|---|
Ep. 8 — Positional Embeddings
How transformers know where words are in a sequence — from sinusoidal encodings to RoPE.
Contents
- Notebook:
positional_embeddings.ipynb
Topics
- Sinusoidal positional encodings (Attention Is All You Need)
- Learned positional embeddings (BERT, GPT-2)
- RoPE — Rotary Position Embedding (Llama, Mistral, Qwen)
- ALiBi — Attention with Linear Biases
- Frequency band analysis and long-context extrapolation
Related
- Downloads last month
- 30