SAE TTS Demo

Single-latent steering examples. (evaluation features)

Example 1 (Latent #3237)

Text: I arrived at the location earlier than expected.

Happy (baseline)

Happy (scaled +60)

Happy (scaled -60)

Example 2 (Latent #24)

Text: I arrived at the location earlier than expected.

Anger (baseline)

Anger (scaled +70)

Anger (scaled -70)

Example 3 (Latent #195)

Text: I observe the results after processing.

Note: Latent #195 (Happy factor.)

Neutral (scaled -50)

Neutral (baseline)

Neutral (scaled +50)

Example 4 (Latent #275)

Text: I arrived at the location earlier than expected.

Sadness (baseline)

Sadness (scaled +100)

Sadness (scaled -100)

Neutral → Target steering (Top6 features)

Example 5 (Neutral → Happy)

Text: The task reqires several steps to complete.

Neutral (baseline)

Happy (scaled +80)

Example 6 (Neutral → Anger)

Text: I arrived at the location earlier than expected.

Neutral (baseline)

Neutral (scaled +60)

Example 7 (Neutral → Sadness)

Text: The process compelete without interruption.

Neutral (baseline)

Sadness (scaled +60)

Target → Neutral steering

Example 8 (Anger → Neutral)

Text: I review the document before sending it.

Anger (baseline)

Neutral (scaled -60)

Example 9 (Sadness → Neutral)

Text:I review the document before sending it.

Note: Latent #4 (Sadness factor.)

Sadness (baseline)

Neutral (scaled -40)

Example 10 (Happiness → Neutral)

Text: The message appeared on the screen.

Happiness (baseline)

Neutral (scaled -50)