steering

# /steering Run activation-steering feature discovery end-to-end: specify a model and an optional feature, and this skill orchestrates data generation, vector extraction, and artifact output. ## When to use this skill - The user types `/steering` or asks to "steer" a model - The user wants to extract a cognitive feature (chain_of_thought, few_shot_prompting, react, etc.) - The user wants to generate steering artifacts for a model - The user asks to run feature discovery or build a steering vector - The user mentions activation engineering, representation engineering, or contrastive extraction ## How it works ### Inputs | Parameter | Default | Description | |-----------|---------|-------------| | `model_name` | `"gpt2"` | Any decoder-only Hugging Face model identifier | | `feature_name` | *auto-pick* | A feature to extract (e.g. `chain_of_thought`). If omitted, the system picks the next undiscovered feature from the standard catalog. | | `user_examples` | *generate* | Optional list of `FeatureExample` objects with `text` and `label` ("positive" / "negative"). If omitted, synthetic examples are generated. If only one label is present, the other is generated. | | `layer_idx` | `5` | Transformer layer for hidden-state collection | | `output_dir` | `None` | Directory to write plugin artifacts to | ### Pipeline 1. **Resolve feature** — look up in the standard catalog, or generate a new spec. 2. **Ensure data** — use user-supplied examples, fill missing labels with synthetic data, or generate all data. 3. **Load model** — `load_model_and_tokenizer(model_name)`. 4. **Discover vectors** — `discover_feature_vectors(feature_spec, layer_idx, model, tokenizer, device)`. 5. **Write artifacts** — `write_artifact_plugin(...)` to produce a distributable plugin bundle. ### Output A `SteeringResult` with: - `feature_spec` — the resolved `FeatureSpec` - `discovered_vectors` — list of `DiscoveredFeatureVector` (torch tensors + metadata) - `artifact_dir` — path to the written plugin directory (if `output_dir` was set) ## Quick-start ```python from activation_steering.steering_command import SteeringRunConfig, run_steering # Explicit feature result = run_steering(SteeringRunConfig( model_name="gpt2", feature_name="chain_of_thought", output_dir="./artifacts", )) # Auto-pick next undiscovered feature result = run_steering(SteeringRunConfig( model_name="gpt2", output_dir="./artifacts", )) # With user-supplied data from activation_steering.features import FeatureExample result = run_steering(SteeringRunConfig( model_name="gpt2", feature_name="my_custom_feature", user_examples=[ FeatureExample(text="detailed reasoning before answer", label="positive"), FeatureExample(text="just the answer", label="negative"), ], output_dir="./artifacts", )) ``` ## Artifacts produced ``` <output_dir>/<model_name>/<feature_name>/ ├── plugin.json # Manifest ├── feature_specs.json # The resolved FeatureSpec └── controllers.json # Discovered steering vectors ``` ## Standard catalog features (gpt2) | Feature | Category | |---------|----------| | `few_shot_prompting` | prompt_engineering | | `retrieval_augmented_context` | context_engineering | | `react` | cognitive_architecture | | `chain_of_thought` | reasoning_strategy | ## Implementation The command is implemented in `activation_steering/steering_command.py` and exported from `activation_steering`: - `SteeringRunConfig` — run configuration dataclass - `SteeringResult` — run output dataclass - `run_steering(config)` — main orchestrator - `build_steering_feature_spec(...)` — resolve or generate a FeatureSpec - `pick_undiscovered_feature(...)` — auto-select next feature - `generate_synthetic_examples(...)` — create training data ## API reference For implementation details, read [`activation_steering/steering_command.py`](../../../activation_steering/steering_command.py).