Skip to main content

Temporal Feature Analysis

Temporal Feature Analysis is a method by Lubana, Rager, Hindupur, et al. It finds features that capture both the context of a text (for example, "this is a story") and local information (for example, "this token is a number").

Start here: Release | Demo | Paper | Notebook

The similarity matrix for a story that changes to Python code. The two bright squares show the two parts of the text.

Similarity Matrix​

The similarity matrix shows how similar the representation of each token is to each other token. Bright squares show parts of the text with the same context.

  1. Open the demo.
  2. Select a demo on the left, or type your own text (at least one sentence is best). Then click Generate.
  3. Hover over a token to highlight its cells in the matrix. Hover over a cell to see its tokens.
  4. Click Share to get a link.

Available Sources​

ModelLayers
Gemma 2 2B12
Llama 3.1 8B15, 26

These sources also have feature dashboards, and you can steer with them.