Steering LLM states on 50 prompts
1.6666LogFreq (d)ActAdd (top/bot 50 logfreq)
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| ActAdd (top/bot 50 logfreq)Model=TinyLlama, Layer=11, Number of prompts=502025.12 | 1.6666 | 0.001 | -0.074 | 0.237 | |
| Brain axis (axis 15)Model=TinyLlama, Layer=11, Number of prompts=502025.12 | 0.9246 | 0.001 | -0.2183 | 0.001 | |
| Random dir. (axis 15)Model=TinyLlama, Layer=11, Number of prompts=502025.12 | -0.02 | 0.88 | — | — |