Steering

Edit the RFD3/RF3 residual stream during inference to push or suppress structural features.

Modes

Mode	Source	Operation
`sae_feature`	`α · W_dec[feature_id]`	Add/ablate a single SAE dictionary feature
`raw_diff`	`coeff · (mean_pos − mean_neg)`	Add/ablate a pre-computed mean-diff vector

Both reduce to the same operation: add a fixed direction to the hook output. Ablation projects the direction out.

Steering config schema


# @package steering
block12:
  - mode: sae_feature           # or raw_diff
    sae_path: outputs/sae/.../final.pt
    feature_id: 639
    alpha: 5.0
    type: steer                 # steer | ablate
    apply_at_steps: all         # all | [10, 20] | {start: 50}

For raw_diff:


block12:
  - mode: raw_diff
    vector_path: outputs/steering/vectors/haz_minus_ben/block12.pt
    coeff: 3.0
    apply_at_steps: all

Recipes

Recipe A — SAE feature


# Single-feature SAE steer along the top hazard direction
# (block12 feature #639, AUROC 0.815, PLA2-correlated)
 
saffron steer model=rfd3 \
  hooks=rfd3_partial \
  steering=sae_block12_f639 \
  inputs=tutorials/steering/configs/steer_block12_f639.json \
  out_dir=outputs/steering/runs/benign_inputs_f639_alpha5

Recipe B — Raw diff


# Step 1 — compute the mean-diff direction
saffron compute_steering_vector \
  inputs=tutorials/steering/configs/diff_haz_minus_ben.yaml \
  out_dir=outputs/steering/vectors/haz_minus_ben
 
# Step 2 — steer
saffron steer model=rfd3 \
  hooks=rfd3_partial \
  steering=raw_diff_block12 \
  inputs=tutorials/steering/configs/steer_diff_block12.json \
  out_dir=outputs/steering/runs/benign_inputs_diff_alpha3

Recipe C — Ablate


# Ablate the hazard direction from hazard inputs
# (does removing the feature de-hazardize the output structure?)
 
saffron steer model=rfd3 \
  hooks=rfd3_partial \
  steering=ablate_block12_f639 \
  inputs=tutorials/steering/configs/ablate_block12_f639.json \
  out_dir=outputs/steering/runs/hazard_inputs_ablate_f639

Null-steer sanity check

α=0 must reproduce saffron collect bit-for-bit. Run this before trusting any steering result.


saffron steer model=rfd3 \
  hooks=rfd3_partial \
  steering=null_block12_f639 \
  inputs=tutorials/steering/configs/null_block12_f639.json \
  out_dir=outputs/steering/runs/null_check

Compute a new direction vector


saffron compute_steering_vector \
  inputs=tutorials/steering/configs/diff_haz_minus_ben.yaml \
  out_dir=outputs/steering/vectors/haz_minus_ben


# diff_haz_minus_ben.yaml
activations_h5: outputs/collect/train/activations/activations.h5
labels_csv: data_pipelines/safeprotein/labels.csv
positive_label: 1
negative_label: 0
hooks: [block6, block8, block12]
normalize: unit
out_dir: outputs/steering/vectors/haz_minus_ben

Outputs: block6.pt, block8.pt, block12.pt, meta.json.

Adding a custom steering config

Create sae/src/sae/configs/steering/my_steer.yaml:


# @package steering
block12:
  - mode: sae_feature
    sae_path: outputs/sae/.../final.pt
    feature_id: 42
    alpha: 3.0
    apply_at_steps: all

Then: saffron steer model=rfd3 hooks=rfd3_partial steering=my_steer inputs=... out_dir=...