Discover optimal steering directions for a task by searching layers and random directions against stored activations, then report the best-scoring configuration.
python -m wisent discover-steering --model MODEL --task TASK --layer N --strength S [OPTIONS]
python -m wisent discover-steering --model meta-llama/Llama-3.1-8B-Instruct --task truthfulqa_custom --layer 15 --strength 1.5 --n-test-samples 100 --n-random-directions 50 --discover-task-limit 500 --discover-train-limit 1000 --output ./discovery.json
python -m wisent discover-steering --model meta-llama/Llama-3.1-8B-Instruct --task truthfulqa_custom --layer 15 --layer-range 8-16 --strength 1.5 --n-test-samples 100 --n-random-directions 50 --discover-task-limit 500 --discover-train-limit 1000 --output ./discovery.json
| Argument | Description |
|---|---|
| --model | Model name |
| --task | Task/benchmark name in database (e.g. truthfulqa_custom) |
| --layer | Primary layer to test |
| --strength | Steering strength multiplier |
| --n-test-samples | Number of test samples for evaluation |
| --n-random-directions | Number of random directions to try |
| --discover-task-limit | Maximum task pairs to load for discovery |
| --discover-train-limit | Maximum activation pairs to load for training |
| --output | Output JSON file path |
| Argument | Default | Description |
|---|---|---|
| --layer-range | only --layer | Layer range to search, e.g. 8-16 |
| --database-url | DATABASE_URL env | Database URL |
| --skip-layer-search | off | Skip the layer search (faster, less thorough) |
| --skip-direction-search | off | Skip the direction search (faster, less thorough) |
Stay in the loop. Never miss out.
Subscribe to our newsletter and unlock Wisent insights.