diff --git a/src/components/Attention.svelte b/src/components/Attention.svelte
index f87540f..11f4047 100644
--- a/src/components/Attention.svelte
+++ b/src/components/Attention.svelte
@@ -139,8 +139,8 @@
{/each}
The model uses the masked self-attention scores and multiplies them with the
- Value matrix to get the
- final output
- of the self-attention mechanism. GPT-2 has 12 self-attention heads, each capturing
- different relationships between tokens. The outputs of these heads are concatenated and passed
- through a linear projection.
+ Value matrix to get each head's attention output. GPT-2 has
+ 12 self-attention heads, each capturing different relationships between tokens.
+ The outputs of these heads are concatenated into one vector and then passed through the
+ attention block's learned output projection matrix (c_proj) before the result
+ flows to the next layer. The projection matrix is part of the model computation, but it is
+ not drawn as a separate matrix in this visualization.
Each head multiplies its attention scores with the Value embeddings to produce its attention output—a refined representation of each token after considering context.
GPT-2 (small) has 12 such outputs, which are concatenated to form a single vector of the original size (768 numbers).
', + 'Each head multiplies its attention scores with the Value embeddings to produce a head-specific attention output—a refined representation of each token after considering context.
GPT-2 (small) concatenates the 12 head outputs, then applies the attention block\'s learned output projection matrix (c_proj) before sending the result to the next layer. This projection is part of the computation even though it is not drawn as a separate matrix here.