Run the Zwiad geometry-analysis protocol with concept decomposition over stored activations: measure signal, geometry, decomposition, intervention and editability, and name discovered concepts with an LLM.
python -m wisent zwiad (--from-database | --from-cache PATH) --model MODEL --steps STEPS [OPTIONS]
python -m wisent zwiad --from-database --model meta-llama/Llama-3.1-8B-Instruct --task truthfulqa_custom --prompt-format chat --extraction-strategy chat_last --extraction-component residual_stream --steps all --llm-model gpt-4o-mini --concept-naming-n-samples 20 --visualizations-dir ./zwiad_viz --output ./zwiad.json
python -m wisent zwiad --from-cache ./activations.pkl --model meta-llama/Llama-3.1-8B-Instruct --prompt-format chat --extraction-strategy chat_last --extraction-component residual_stream --steps signal,geometry --llm-model gpt-4o-mini --concept-naming-n-samples 20 --visualizations-dir ./zwiad_viz
| Argument | Description |
|---|---|
| --from-database | Load activations from the Supabase database |
| --from-cache PATH | Load activations from a local cache pickle file |
| Argument | Description |
|---|---|
| --model | Model name in database |
| --prompt-format | Prompt format used for activations (chat or completion) |
| --extraction-strategy | Token extraction strategy (last_token, first_token, chat_last) |
| --extraction-component | Transformer component to extract from (residual_stream, attn_output, mlp_output, ...) |
| --steps | Protocol steps: all, signal, geometry, decomposition, intervention, editability, or comma-separated |
| --llm-model | LLM model for concept naming |
| --concept-naming-n-samples | Samples used for LLM concept-naming prompts |
| --visualizations-dir | Directory to save visualization PNGs |
| Argument | Default | Description |
|---|---|---|
| --task | (none) | Task name (required for --from-database; enables concept naming for --from-cache) |
| --layers | all | Layers to analyze: range (8-16) or comma-separated (8,10,12) |
| --output | (none) | Output file path for JSON results |
| --database-url | DATABASE_URL env | Database URL |
| --no-visualizations | off | Disable visualization generation |
The geometry metrics expose additional design-choice tuning flags (probe, direction-stability, consistency, sparsity and detection thresholds); run wisent zwiad --help for the full list.
Stay in the loop. Never miss out.
Subscribe to our newsletter and unlock Wisent insights.