license: mit
tags:

• interpretability
• ai-safety
• chain-of-thought
• steering-vectors
• activation-steering
  pretty_name: CoT-controllability steering vectors (gpt-oss-20b)

CoT-controllability steering vectors — artifacts

Artifacts for the project "A 2,880-number steering vector gives a reasoning model the
chain-of-thought control that fine-tuning does" on gpt-oss-20b. Code + the master notebook +
generate_figures.py that load these artifacts:
https://github.com/redwoodresearch/automated-research-projects (folder
cot-controllability-steering-vectors).

Contents

• steering_vectors/ — the headline frozen-weights steering vector grad_steer_gL10.npz
  (layer 10, 2,880 floats, ‖v‖≈148) plus the full family (seeds, layers gL6/gL8/gL12, control twin,
  sign-reversed, random, multi-layer; the faithfulness/concealment vectors; the reliance vectors) and
  the diff-of-means directions (steering_directions.npz, ftbase_direction.npz). Each .npz stores
  layers + vec_<i>; a sibling *_meta.json records the training provenance.
• datasets/ — the novel datasets: the source-stratified task pool (tasks_all*.jsonl), the
  instruction splits (instruction_splits.json; 25 instructions / 6 categories, formatting category +
  bullet probe held out), the edited-reasoning SFT data + the raw-trace control, the natural source
  traces, the plain mix, the cue/faithfulness/reliance target sets, and the fresh GSM-Symbolic tasks.
• figure_data/ — the small summary JSONs generate_figures.py / the notebook plot for fig1–fig5.
• results_raw/ — the raw artifacts needed to re-derive figure_data/ (the per-example fig5
  attention tensors tok_subspan_attn.npz, the per-row judged generations behind fig1/fig2) plus the
  headline judged eval files. The full ~700 MB of per-(task×instruction) generations is available on
  request.

How the figures use these

generate_figures.py --source hf downloads figure_data/*.json from this repo and regenerates
fig1–fig5 on CPU (no model, no GPU). precompute_figure_data.py re-derives figure_data/ from
results_raw/. The steering vectors are loaded with cot_steering.steering.load_steering_vector.

All results are on a single model (gpt-oss-20b); see the GitHub README for scope/limitations.